Early Deliverables

1. A report describing (social) web single sign-on and their security problems and explaining which of these security problems can be detected without knowledge of the inner workings of the web services.

Final Deliverables

1. A web crawler that detects insecure usage of (social) web single sign-on based on OpenID Connect 1.0 and OAuth 2.0.
2. Complete report

Reading

• https://tools.ietf.org/html/rfc6749 (OAuth 2.0)
• https://tools.ietf.org/html/rfc6750 (Bearer Token Usage)
• https://tools.ietf.org/html/rfc6819 (OAuth 2.0 Threat Model and Security Considerations)
• https://tools.ietf.org/html/draft-ietf-oauth-security-topics-21 (Draft OAuth Security BCP)
• https://tools.ietf.org/html/draft-ietf-oauth-browser-based-apps-12 (Draft Best Practices for OAuth Browser-Based Apps)
• https://openid.net/specs/openid-connect-core-1_0.html (OpenID Connect Core 1.0)
• Ronghai Yang, Guanchen Li, Wing Cheong Lau, Kehuan Zhang, and Pili Hu. 2016. Model-based Security Testing: An Empirical Study on OAuth 2.0 Implementations. ASIA CCS '16. ACM, New York, 651–662. DOI:https://doi.org/10.1145/2897845.2897874
• Shernan E., Carter H., Tian D., Traynor P., Butler K. 2015. More Guidelines Than Rules: CSRF Vulnerabilities from Noncompliant OAuth 2.0 Implementations. In: Almgren M., Gulisano V., Maggi F. (eds) Detection of Intrusions and Malware, and Vulnerability Assessment. DIMVA 2015. Lecture Notes in Computer Science, vol 9148. Springer, Cham. https://doi.org/10.1007/978-3-319-20550-2_13
• San-Tsai Sun and Konstantin Beznosov. 2012. The devil is in the (implementation) details: an empirical analysis of OAuth SSO systems. CCS '12. ACM, New York, 378–390. DOI:https://doi.org/10.1145/2382196.2382238 - Yuchen Zhou and David Evans. 2014. SSOScan: Automated Testing of Web Applications for Single Sign-On Vulnerabilities. USENIX Security 2014. https://www.usenix.org/conference/usenixsecurity14/technical-sessions/presentation/zhou
• Daniel Fett, Ralf Küsters, Guido Schmitz. 2016. A Comprehensive Formal Security Analysis of OAuth 2.0. In: 23rd ACM SIGSAC Conference on Computer and Communications Security (CCS 2016). https://dl.acm.org/doi/pdf/10.1145/2976749.2978385
• Daniel Fett, Ralf Küsters, Guido Schmitz. 2017. The Web SSO Standard OpenID Connect: In-Depth Formal Security Analysis and Security Guidelines. In: IEEE 30th Computer Security Foundations Symposium (CSF 2017). https://ieeexplore.ieee.org/iel7/8048777/8049639/08049720.pdf

Early Deliverables

1. A report describing current approaches to integrate apps into MS Teams and Slack
2. A report with an overview of possible security services to integrate in the chatbot (VirusTotal, etc.)
3. A proof of concept app for either MS Teams and Slack (following the tutorials would be enough)

Final Deliverables

1. A chatbot that allows users to access the functionality of a specific API via MS Teams or Slack (the bot may include a setup phase were API keys need to be provided)
2. A test suite to verify that all implemented tests work as intended
3. A security analysis of the chatbot app
4. A user manual for the bot
5. A critical comparison with related works

Reading

• https://api.slack.com
• https://docs.microsoft.com/en-us/microsoftteams/platform/
• https://link.springer.com/chapter/10.1007/978-3-319-99927-2_3
• https://s3lab-rhul.slack.com/apps/A81FQ3116-1password

Early Deliverables

1. A report on hardware RNG and appropriate third-party secure modules (such as TPM), including the methods used by manufacturers to test randomness (and how end users are extended the opportunity to perform some tests for themselves).
2. A report on the current state of the art in supply chain attacks against critical security infrastructure.
3. A clear workplan describing the set of tests to be performed, including any simulation, emulation or experimental systems that will be used.

Final Deliverables

1. A simulation or experimental demonstration of structured information that can pass simple statistical tests of randomness.
2. Appropriate remediation measures, be they existing or novel tests of randomness, with appropriate justifications (explaining why the additional time, cost or other resources are warranted and under what circumstances).
3. A final report documenting the project. This report will extend to cover the experimental/simulation design, generation of biased random sequences, and how lightweight statistical tests were selected and evaluated. A clear grasp of experimental process and hypothesis testing should be demonstrated through appropriate use of tools and techniques.

Reading

• https://ieeexplore.ieee.org/abstract/document/9069949/
• https://link.springer.com/chapter/10.1007/978-3-030-04762-7_8
• https://dl.acm.org/doi/abs/10.1145/3264628
• https://dl.acm.org/doi/abs/10.1145/1268776.1268777
• https://ieeexplore.ieee.org/abstract/document/6135498/

Early Deliverables

1. Proof of concept program: A prototype implementation of a board game (e.g., chess).
2. Proof of concept program: A prototype program that exhibits the behaviour of concurrent execution.
3. Report: A survey report on game environments.
4. Report: A description of the prototype implementations.

Final Deliverables

1. The system must be implemented according to modern software engineering principles.
2. The environment should contain a working system, with functionalities such as error reporting.
3. A graphical interface should be implemented.
4. The environment should allow the users to play multiple games at the same time.
5. The report should provide an overview of game-playing environments.
6. The report should describe the environment including its functionalities and design.
7. The report should describe the implementation issues (such as the choice of data structures, numerical methods etc) necessary to apply the theory.
8. The report should describe the testing procedures of the environment.
9. The report should describe the software engineering process involved in the design and implementation.

Suggested Extensions

• Advanced GUI with other useful features such as chat room.
• Advanced implementation issues such as multiple game playing based on concurrency and sharing.

Early Deliverables

1. Proof of concept program: A prototype implementation of a board game (e.g., chess).
2. Proof of concept program: A prototype program that exhibits the behaviour of concurrent execution.
3. Report: A survey report on game environments.
4. Report: A description of the prototype implementations.

Final Deliverables

1. The system must be implemented according to modern software engineering principles.
2. The environment should contain a working system, with functionalities such as error reporting.
3. A graphical interface should be implemented.
4. The environment should allow the users to play multiple games at the same time.
5. The report should provide an overview of game-playing environments.
6. The report should describe the environment including its functionalities and design.
7. The report should describe the implementation issues (such as the choice of data structures, numerical methods etc) necessary to apply the theory.
8. The report should describe the testing procedures of the environment.
9. The report should describe the software engineering process involved in the design and implementation.

Suggested Extensions

• Advanced GUI with other useful features such as chat room.
• Advanced implementation issues such as multiple game playing based on concurrency and sharing.

Early Deliverables

1. A text-based (non-interactive) monochrome web-page
2. A colourful web-site including images and navigation
3. GUI built with buttons etc.
4. Report: about 15 pages including sketches of designs.

Final Deliverables

1. Design and implement a more advanced interface(s)
2. Complete report
3. The programs must have an object-oriented design, using modern software engineering principles.
4. The report will describe the software engineering processes involved in generating your software.
5. The report will include comparisons of interfaces with a discussion of their meanings.
6. The report will include a User Manual.

Reading

• http://hci.rwth-aachen.de/HCIBooks
• http://www.netmagazine.com/features/top-50-books-web-designers-and-developers

Early Deliverables

1. A text-based (non-interactive) monochrome web-page
2. A colourful web-site including images and navigation
3. GUI built with buttons etc.
4. Report: about 15 pages including sketches of designs.

Final Deliverables

1. Design and implement a more advanced interface(s)
2. Complete report
3. The programs must have an object-oriented design, using modern software engineering principles.
4. The report will describe the software engineering processes involved in generating your software.
5. The report will include comparisons of interfaces with a discussion of their meanings.
6. The report will include a User Manual.

Reading

• http://hci.rwth-aachen.de/HCIBooks
• http://www.netmagazine.com/features/top-50-books-web-designers-and-developers

Early Deliverables

1. A report describing techniques/methods to create malicious code that look benign
2. Techniques (e.g., AI-based) that can be used to generate malicious code autonomously
3. A clear workplan describing the design of the system to be implemented and set of tests to be performed to verify the generated code

Final Deliverables

1. A thorough analysis of examples of automatically-generated malicious code that pass rigorous inspection
2. Describe in detail the design of the system and of the tests
3. Describe and analyse the (AI-based) techniques and features that have been used during the project to automatically generate malicious code
4. Critically compare the findings with related works
5. Suggestions for improving the inspection of code generated by the proposed tool

Reading

• https://en.wikipedia.org/wiki/Underhanded_C_Contest
• http://underhanded-c.org/

Early Deliverables

1. A report describing the state-of-the-art of Web development, including technologies/frameworks/platforms and their comparison as described above. The report will also explain the concrete choices made and the respective rationale/justification from a software engineering perspective (i.e. a justification merely based on a skills learning perspective will not suffice)
2. A report on Web development architectural paradigms and applicable design patterns, including an initial discussion of security, privacy and key-operational aspects and considerations (e.g. cloud deployments and DevOps aspects)
3. A report describing the system to be implemented, including a comprehensive set of well-defined use cases (or user stories).
4. A functional prototype implementation of the system, clearly demonstrating sound design principles and implementation practices (e.g. on DB design, on UI design and interaction).

Final Deliverables

1. A fully functional software application, with an appropriate end-to-end architecture, applicable design patterns, using modern Web technologies and rigorous software engineering principles.
2. Complete report, encompassing also relevant content (further explained/discussed) from the early deliverables reports.
3. The report will describe the software engineering processes involved in generating your software.
4. The report will include a User Manual.

Early Deliverables

1. A simple proof-of-concept implementation of the aggregating algorithm and fixed share algorithm.
2. Report: An overview of the problem of prediction with expert advice and the work of the aggregating algorithm.
3. Report: The performance of the aggregating algorithm on a small artificial dataset.

Final Deliverables

1. The program will work with a real dataset, read the data from a file, preprocess it, and use a standard algorithm to generate experts' predictions (if necessary).
2. The program will implement aggregating algorithm plus possibly fixed share and the aggregating algorithm for sleeping experts. The results will be visualised.
3. The report will describe implementation issues (such as the choice of data structures, numerical methods etc) necessary to apply the theory.
4. The report will describe computational experiments, analyse their results and draw conclusions.

Suggested Extensions

• Application of sleeping experts for predicting the implied volatility of options.
• Experiments with the weak aggregating algorithm.

Reading

• Y.Kalnishkan, The AA And Laissez-Faire Investment http://onlineprediction.net/?n=Main.TheAAAndLaissez-FaireInvestment
• Kalnishkan, Y. Prediction with expert advice for a finite number of experts: A practical introduction. Pattern Recognition, 126 (2022).
• Vovk, V. Competitive on‐line statistics. International Statistical Review, 69(2), 213-248 (2001).
• M.Herbster and M.Warmuth, Tracking the Best Expert, Machine Learning, 32, 151-178 (1998)
• Y.Kalnishkan and M.V.Vyugin. The weak aggregating algorithm and weak mixability. Journal of Computer and System Sciences, 74(8): 1228--1244 (2008)

Early Deliverables

1. A simple proof-of-concept implementation of the aggregating algorithm and fixed share algorithm.
2. Report: An overview of the problem of prediction with expert advice and the work of the aggregating algorithm.
3. Report: The performance of the aggregating algorithm on a small artificial dataset.

Final Deliverables

1. The program will work with a real dataset, read the data from a file, preprocess it, and use a standard algorithm to generate experts' predictions (if necessary).
2. The program will implement aggregating algorithm plus possibly fixed share and the aggregating algorithm for sleeping experts. The results will be visualised.
3. The report will describe implementation issues (such as the choice of data structures, numerical methods etc) necessary to apply the theory.
4. The report will describe computational experiments, analyse their results and draw conclusions.

Suggested Extensions

• Application of sleeping experts for predicting the implied volatility of options.
• Experiments with the weak aggregating algorithm.

Reading

• Y.Kalnishkan, The AA And Laissez-Faire Investment http://onlineprediction.net/?n=Main.TheAAAndLaissez-FaireInvestment
• Kalnishkan, Y. Prediction with expert advice for a finite number of experts: A practical introduction. Pattern Recognition, 126 (2022).
• Vovk, V. Competitive on‐line statistics. International Statistical Review, 69(2), 213-248 (2001).
• M.Herbster and M.Warmuth, Tracking the Best Expert, Machine Learning, 32, 151-178 (1998)
• Y.Kalnishkan and M.V.Vyugin. The weak aggregating algorithm and weak mixability. Journal of Computer and System Sciences, 74(8): 1228--1244 (2008)

Early Deliverables

1. Report: A general report on algorithms for economics social choice, finishing with a determination of the project topic or topics.
2. Report: A full formal description of the problem and the solution concept.
3. Report: Known algorithms on the topic. How they work?
4. Proof of concept: A program that generates a (small) instance of the problem(s) you are studying and checks whether a given solution is indeed correct.
5. Proof of concept: Simple greedy heuristics for the topic.
6. Implementation of a known algorithm on the topic that works on some toy-examples of the problem.

Final Deliverables

1. A program developped with proper software engineering techniques
2. A nice and easy to use GUI for the chosen topic that allows interesting experiments (the GUI and its functionalities depend on the topic).
3. The program should work on at least three special cases of the problem.
4. Implementation and comparison of different heuristics and standard algorithms.
5. Final report should contain a section on the outcomes of the experiments done and conclusions drawn about which algorithms seem suitable in different situations.

Suggested Extensions

• A more careful theory report, including notions of problem hardness, or running time for the algorithms.
• More involved algorithms for the problem.
• In depth analysis of the algorithms for specific classes of instances.
• Nice analysis options for your problem including an implementation in some Discrete Event Simulation framework
• An interactive Web page allowing people to explore and understand your topic using examples and simulation.

Reading

• The table of contents of the Handbook of Computational Social Choice is very useful to see what kinds of project you might want to do
• This course on Algorithmic Game Theory is worth having a skim over. At least the list of titles.
• There is also a good book on Algorithmic Game Theory
• Nash equilibrium: The Nash Equilibrium Wiki page is nice as an introduction.
• Stable Marriage: The Stable Marriage Wiki page is also nice as an introduction.
• Some fun Youtube videos about matching - worth watching even if you do not take this project: video one, or video two.
• A more technical introduction to Networks of Complements.
• A more technical introduction to Schelling games.
• A more technical introduction to Coordination Games on Graphs.
• A more technical introduction to Fair Allocation of Indivisible Goods.
• A (slightly) more technical introduction to Cake cutting algorithms.
• A more technical introduction to Matching Markets

Early Deliverables

1. Report: A general report on algorithms for economics social choice, finishing with a determination of the project topic or topics.
2. Report: A full formal description of the problem and the solution concept.
3. Report: Known algorithms on the topic. How they work?
4. Proof of concept: A program that generates a (small) instance of the problem(s) you are studying and checks whether a given solution is indeed correct.
5. Proof of concept: Simple greedy heuristics for the topic.
6. Implementation of a known algorithm on the topic that works on some toy-examples of the problem.

Final Deliverables

1. A program developped with proper software engineering techniques
2. A nice and easy to use GUI for the chosen topic that allows interesting experiments (the GUI and its functionalities depend on the topic).
3. The program should work on at least three special cases of the problem.
4. Implementation and comparison of different heuristics and standard algorithms.
5. Final report should contain a section on the outcomes of the experiments done and conclusions drawn about which algorithms seem suitable in different situations.

Suggested Extensions

• A more careful theory report, including notions of problem hardness, or running time for the algorithms.
• More involved algorithms for the problem.
• In depth analysis of the algorithms for specific classes of instances.
• Nice analysis options for your problem including an implementation in some Discrete Event Simulation framework
• An interactive Web page allowing people to explore and understand your topic using examples and simulation.

Reading

• The table of contents of the Handbook of Computational Social Choice is very useful to see what kinds of project you might want to do
• This course on Algorithmic Game Theory is worth having a skim over. At least the list of titles.
• There is also a good book on Algorithmic Game Theory
• Nash equilibrium: The Nash Equilibrium Wiki page is nice as an introduction.
• Stable Marriage: The Stable Marriage Wiki page is also nice as an introduction.
• Some fun Youtube videos about matching - worth watching even if you do not take this project: video one, or video two.
• A more technical introduction to Networks of Complements.
• A more technical introduction to Schelling games.
• A more technical introduction to Coordination Games on Graphs.
• A more technical introduction to Fair Allocation of Indivisible Goods.
• A (slightly) more technical introduction to Cake cutting algorithms.
• A more technical introduction to Matching Markets

Early Deliverables

1. Report: A full description of the facility location problem
2. Report: P vs NP, NP-completeness, NP-hardness
3. Report: Strategies for solving NP-complete problems (ex: heuristics, approximations, algorithms for special cases and non-worst case inputs).
4. Report: Linear programming, LP-relaxations and LP Software packages
5. Worker threads in a GUI?
6. Proof of concept program: A simple greedy algorithm
7. Proof of concept program: A program which solves the LP-relaxation of the problem (by using an LP-solver software package)

Final Deliverables

1. The program should be able to read a problem description from an input file, in one of the formats used in popular benchmark libraries
2. The program should have a GUI, which displays a problem instance and can be used to run several different algorithms to solve the instance
3. The program should support several different algorithms to solve a problem, including at least one heuristic (e.g., a greedy algorithm) and at least one algorithm using the LP-solution as guidance. (It is recommended that one of the algorithms comes from the literature on approximation algorithms; see Extensions.)
4. The program should also provide the option of showing the cost of the LP solution, and should ideally provide the option of attempting to produce an exact solution via integer programming (so-called MIP).
5. Since some of the computations can be lengthy (e.g., the MIP solver), the program must allow the user a way to abort an algorithm. Therefore the program must be threaded, so that the user can interact with the GUI despite some algorithm being busy in the background.

Suggested Extensions

• A more careful treatment of approximation: A report on "Approximation algorithms and approximation ratios", and an implemention of various types of approximation algorithm
• More algorithms and solvers. Examples include such meta-heuristics as local search, simulated annealing, etc.)
• Reports on problem variants and extensions (examples: capacitated, multi-level, or the k-center/k-median problem), perhaps including motivations and theoretical properties (NP-hardness, approximation properties)
• Careful experiments with running times and result quality for different algorithms and problem instances

Early Deliverables

1. Report on connected components in undirected graphs and and strongly connected components in directed graphs
2. Report on distances in directed graphs.
3. Some pseudo-code for important algorithms
4. Implementation and testing of some important algorithms.

Final Deliverables

1. Reports for a wider selection of graph algorithms
2. Extensive and accurate commented Pseudo-code
3. Full implementation and testing of algorithms.

Suggested Extensions

• Nice User friendly GUI for algorithms.

Early Deliverables

1. Proof of concept programs: parsers for biological databases,
2. Proof of concept programs: implementation of one clustering algorithm,
3. Proof of concept programs: design of an interface.
4. Reports: Biological databases.
5. Reports: Clustering algorithms.
6. Reports: Design Patterns in Java (mainly plugin architecture).

Final Deliverables

1. The software should be able to work with at least two kind of biological data (microarray data and sequence data) and two clustering algorithms (for instance k-means and hierarchical clustering).
2. At least one visualisation/assessment technique should be added (for example, over-representation analysis).
3. The report should include a biological overview of the problem,
4. The report should include a detailed description of the implementation of the plugin-architecture,
5. The report should include a comprehensive documentation on how to extend the program.

Suggested Extensions

• Adding the possibility of processing more biological data (PPI data, for example, CSV, etc)
• Adding more clustering algorithms, possibly using the Façade pattern with existing implementations (SCSP, Cluster ONE, ...).
• Developing cluster visualisation techniques: MDS visualisation, dendrograms, etc.

Early Deliverables

1. Introduction to Horn clauses and their analysis in Z3.
2. Exposition of the encoding of pushdown systems as Horn-clauses (based on a summary algorithm).
3. Implementation of pushdown analysis by reduction to Horn-clauses.

Final Deliverables

1. Encoding of the CSS redundancy problem for the simplified tree language as Horn-clauses.
2. Implementation of the above encoding.
3. Benchmarking of implementation on examples derived from jQuery code.
4. Automatic, implemented translation from some fragment of HTML and jQuery to the CSS redundancy problem for the simplified tree language.

Early Deliverables

1. A report describing client honeypots and their use.
2. [Optional] Set up a client honeypot system and gather data.
3. A report on client honeypot data, gathering, processing, enhancing (e.g. gathering other data from VirusTotal etc).
4. Design of a visualisation, monitoring and data analysis system.

Final Deliverables

1. Design and develop proof of concept for some or all of a visualisation, monitoring and data analysis system. An example might be a program for extracting features from honeypot log file traces and clustering the resulting data.
2. Develop an experiment to gather data and answer a research question e.g. to determine whether phishing sites use drive-by-downloads, or to determine whether a sample of current twitter URLs are malicious, etc.
3. Data analysis and report, e.g. create open source intelligence or show clustering of data and justification of the machine learning methods used and example outputs.
4. Final report including critical comparison, evaluation and visualisation of results showing the capabilities of the tool using different selected features and techniques.

Reading

• [1] Pattern Recognition and Machine Learning, by C.M. Bishop, Springer 2006
• [2] Machine learning: a probabilistic perspective, by K. Murphy, MIT Press 2012
• [3] Virtual Honeypots: From Botnet Tracking to Intrusion Detection Paperback, by Neils Provos and Thorsten Holz, Wiley, 2007
• [4] Seifert, Christian, Ramon Steenson, Ian Welch, Peter Komisarczuk, and Barbara Endicott-Popovsky. "Capture–A behavioral analysis tool for applications and documents." digital investigation 4 (2007): 23-30. http://dfrws.org/sites/default/files/session-files/paper-capture_-_a_tool_for_behavioral_analysis_of_applications_and_documents.pdf
• [5] Dell'Aera, Angelo. "Thug: a new low-interaction honeyclient." Recuperado el 11 (2012). http://www.tlc.unipr.it/veltri/sicurezza-2011-2012/Sec-4-Threats-03_Honeyclient.pdf
• [6] https//github.com/buffer/thug

Early Deliverables

1. Study conformal prediction theory and meaning of its validity and efficiency properties.
2. Pick artificial or natural data examples with various level of difference between the data segments.
3. Implement Conformal Prediction with few versions of conformity score functions and compare their performance on these data examples.

Final Deliverables

1. Survey of conformity scores for CP framework based on standard machine learning algorithms.
2. For one of the algorithms, develop and compare different ways to define conformity score function.
3. Compare their validity and efficiency on the data examples including the transfer learning challenge.

Suggested Extensions

• To try the same for other underlying algorithms, and to formulate general principles behinds safe conformity score functions

Reading

• Hedging predictions in machine learning by Alex Gammerman and Vladimir Vovk. Computer Journal 50:151--177 (2007).
• A tutorial on conformal prediction by Glenn Shafer and Vladimir Vovk. Journal of Machine Learning Research 9:371--421 (2008).

Early Deliverables

1. Study conformal prediction theory and meaning of its validity and efficiency properties.
2. Pick artificial or natural data examples with various level of difference between the data segments.
3. Implement Conformal Prediction with few versions of conformity score functions and compare their performance on these data examples.

Final Deliverables

1. Survey of conformity scores for CP framework based on standard machine learning algorithms.
2. For one of the algorithms, develop and compare different ways to define conformity score function.
3. Compare their validity and efficiency on the data examples including the transfer learning challenge.

Suggested Extensions

• To try the same for other underlying algorithms, and to formulate general principles behinds safe conformity score functions

Reading

• Hedging predictions in machine learning by Alex Gammerman and Vladimir Vovk. Computer Journal 50:151--177 (2007).
• A tutorial on conformal prediction by Glenn Shafer and Vladimir Vovk. Journal of Machine Learning Research 9:371--421 (2008).

Early Deliverables

1. Project plan with tasks and milestones.
2. An overview of Description Logic that supports ontological reasoning.
3. Gathering of requirements from a real user, or justifying such requirements from the perspective of a potential user.
4. Initial experiments with Protege. And building a preliminary model for the chosen domain.

Final Deliverables

1. An ontology of the knowledge domain associated with the data set.
2. A discussion of Description Logic with explicit examples in the ontology constructed.
3. Investigate inferencing algorithms in Description Logic and their run-time complexity.
4. At least a partial encoding of the data set in the ontology.
5. The report will include examples of usage of the system.
6. The report will include discuss the perceived advantages and disadvantages of using ontology frameworks for representing and querying the chosen data set

Suggested Extensions

• Connect the ontology to other existing ontologies in related areas.
• User-friendly interfaces for data representation and querying.

Reading

• Semantic Web for the Working Ontologist. Dean Allemang and Jim Hendler.
• A Description Logic Primer. Markus Krotzsch, Frantiek Simancik, Ian Horrocks.
• http://protege.stanford.edu

Early Deliverables

1. Familiarisation with the modelling languages for automated planning. Show examples of models that describe specific applications and validate them.
2. Familiarisation with state-of-the-art planning systems. Show manually worked examples of their operation.
3. For modelling projects, identification of a specific scenario/application to model as a planning task. In depth description of why planning is well suited to model such a scenario/application. Early prototypes of the models and their validation. For algorithm-oriented projects, identification of a specific planning algorithm to improve. In depth description of how the algorithm can be improved. Early prototypes of improved algorithm with relating experiments.

Final Deliverables

1. For modelling projects, planning model for the chosen application as well as in depth report that describes and motivates the model. Experimental evidence that the problem under consideration can be solved efficiently via planning. Code developed in order to adapt existing planning algorithms to solve the problem under consideration efficiently.
2. For algorithm-oriented projects, code of the algorithms implemented, experimental evidence that the algorithms work efficiently and in depth report on the theoretical approach behind the algorithms implemented.

Reading

• Artificial Intelligence: A Modern Approach (Third edition) by Stuart Russell and Peter Norvig. Chapters 10 and 11.
• A Concise Introduction to Models and Methods for Automated Planning by H. Geffner and B. Bonet. Morgan and Claypool, 2013
• Automated Planning - Theory and Practice by Malik Ghallab, Dana S. Nau, Paolo Traverso. Elsevier, 2004

Early Deliverables

1. A report describing techniques/methods to extract features from binaries
2. A preliminary tool to extract features from binaries (e.g., control-flow graphs)
3. Creation of a reasonably large dataset composed of malware binaries coming from different authors or groups (e.g., APT)
4. A preliminary set of tests with machine learning classifiers (using scikit-learn) to classify a subset of the dataset based on the extracted features

Final Deliverables

1. A thorough analysis of techniques and methods to extract features from binaries
2. Describe and analyse the techniques and features that have been used during the project to extract and attribute binaries
3. Describe in detail the design of the attribution system
4. Test the system on the entire dataset and discuss the results
5. Suggestions for improving the proposed tool

Reading

• https://www.sciencedirect.com/science/article/pii/S1742287614000176
• https://arxiv.org/abs/1512.08546
• https://www.usenix.org/conference/usenixsecurity15/technical-sessions/presentation/caliskan-islam
• https://dl.acm.org/doi/abs/10.1145/3292577
• https://link.springer.com/chapter/10.1007/978-3-642-23822-2_10

Early Deliverables

1. A report describing previous literature in privacy policy analisys
2. A report describing libraries for natural language processing
3. A report describing tools for static analysis of Android apps

Final Deliverables

1. A tool that uses NLP to perform analysis of Android app privacy policies
2. A detailed description of the tests and evaluation performed on the tool
3. A detailed account of the results obtained and an analysis of them
4. A critical comparison with findings from related works

Reading

• https://www.nltk.org
• https://spacy.io
• https://www.usenix.org/conference/usenixsecurity15/technical-sessions/presentation/wang-ruowen
• https://www.usenix.org/conference/usenixsecurity18/presentation/harkous
• https://www.usenix.org/conference/usenixsecurity14/technical-sessions/presentation/zimmeck
• https://frida.re
• https://github.com/skylot/jadx

Early Deliverables

1. Experiments: Set up of an existing symbolic execution engine and experiments with example code
2. Code: Test infrastructure for symbolically executing the target interpreter
3. Report: Mechanics of symbolic execution and state of the art
4. Report: Comparison of two symbolic engines and the results from running them on the interpreter
5. Report: Overview of the implementation of the target interpreter

Final Deliverables

1. Code: A mature testing system and performance optimisations for symbolically executing an interpreted language
2. Report: An evaluation of its efficiency and effectiveness on a small set of benchmarks
3. Report: Process for developing the new engine, including interpreter modifications
4. Report: Testing infrastructure and libraries for the target language
5. Report: Responsible disclosure of vulnerabilities
6. Development and evaluation of high-level assertions of program correctness that allow to find concrete vulnerabilities and classes of bugs relevant to the target programs under test and/or the entire target interpreted language.

Suggested Extensions

• Novel heuristics for improving performance and test coverage.
• Detailed evaluation of the effects of different combinations of optimisations.

Reading

• Bucur et al.: Prototyping Symbolic Execution Engines for Interpreted Languages. Intl. Conference on Architectural Support for Programming Languages and Operating Systems (ASPLOS), Salt Lake City, UT, March 2014
• Cadar and Sen: Symbolic execution for software testing: three decades later. Communications of the ACM, 56:2, February 2013

Early Deliverables

1. A report describing possible applications of homomorphic encryption
2. A report describing one homomorphic encryption scheme
3. A performance comparison of two homomorphic encryption libraries implementing that scheme for a simple homomorphic evaluation, e.g. weighted average on a small dataset

Final Deliverables

1. Describe commonly-used homomorphic encryption schemes and their implementations in open-source libraries
2. Explain how scheme / libraries parameters should be set to optimise performance
3. Compare performance of two libraries for a more involved evaluation, using one or more advanced techniques such as batching or bootstrapping
4. Based on implementation experience, evaluate suitability of different libraries for a given application

Reading

• https://homomorphicencryption.org
• http://www.humangenomeprivacy.org/2019/
• https://github.com/Microsoft/SEAL
• https://github.com/homenc/HElib
• https://eprint.iacr.org/2016/421
• https://eprint.iacr.org/2012/144

Early Deliverables

1. A report giving an overview of prominent distributed multi-shot consensus algorithms.
2. A report describing consensus algorithms in the context of prominent permissioned and/or permissionless Blockchains.
3. A proof-of-concept program implementing a simple (possibly) inefficient multi-shot consensus algorithm

Final Deliverables

1. A permissioned Blockchain that is resilient to crash and/or Byzantine failures and achieves a high transaction throughput.
2. A report describing the architecture of the implemented Blockchain
3. A report on the performed experiments and performance evaluation results

Suggested Extensions

• Support for nodes joining and leaving the system.
• An integration with the Jepsen distributed testing framework.

Reading

• Paxos made simple, L. Lamport, ACM SIGACT News, December 2001, https://www.microsoft.com/en-us/research/publication/paxos-made-simple
• The Part-Time parliament, L. Lamport, ACM TOCS '98: https://www.microsoft.com/en-us/research/publication/part-time-parliament/
• Lecture notes on Leader-based Sequence Paxos - an understandable sequence consensus algorithm, S. Haridi et al, https://arxiv.org/pdf/2008.13456.pdf
• In search of an understandable consensus algorithm, D. Ongaro et al, USENIX ATC '14 https://www.usenix.org/conference/atc14/technical-sessions/presentation/ongaro
• Streamlet: Textbook Streamlined Blockchains, B. Chan and E. Shi, https://eprint.iacr.org/2020/088
• Practical Byzantine fault-tolerance, M. Castro and B. Liskov, OSDI '99: http://pmg.csail.mit.edu/papers/osdi99.pdf

Early Deliverables

1. A simple program displaying an animation using sprites or similar animation technology.
2. A program displaying a simple, interactive user interface. You might want to explore the different means of displaying and writing user interfaces (XML, Object Orientated, Immediate Mode).
3. 3D demo using an API
4. A program that exhibits some 3D graphics using a relatively low level graphics API (Vulkan, OpenGL, WebGL, JavaFX etc. This might provide a starting point for choosing technology: https://en.wikipedia.org/wiki/List_of_3D_graphics_libraries)
5. A simple game using a game engine / library / framework. Think of it as an opportunity to learn about the technologies you'll use for your main game but keep it simple.
6. A report on design patterns in games, an enumeration of commons ones, the problems they solve and perhaps discuss some specific instances of their use.
7. A report on the specific technologies you've tried / intend to use for your game. You should do deep on you intended tech, show as deep an understanding as you can.
8. A report on animation techniques, enumerate them, perhaps tell a history of them, compare and contrast. The goal is to show a deep understanding of how animation in games is achieved.
9. A report on Threading an User Interfaces. What problems do games experience with user interfaces without threading, how does multi threading solve these problems and what different ways can this parallelism or concurrency be achieved? Perhaps review how it has actually been achieve in technologies or platform relevant to your game.
10. A report on designing games with graphical user interfaces for multiple platforms.
11. A report on the technologies used to write and draw graphical user interfaces.

Final Deliverables

1. Game Source Code: You should use a consistent architectural style. Our suggestion is object orientation but you could opt for another with sufficient justification.
2. Game Source Code: The game needs to have at least 3 'screens' or pages that are interactive using the GUI technology of choice.
3. Game Source Code: You should publish your game! (Android Store, itch.io)
4. Game Source Code: The game play should make use of animation.
5. The Final Report: An in depth review of background materials you used to learn the concepts and technologies used to write the game, perhaps presenting a historical timeline of those concepts / technologies.
6. The Final Report: A description of the more difficult concepts involved in writing graphical games e.g. Multi Threading and threads / processes, event based systems (handlers, events), special consideration for the environment your game with run in (memory foot print, battery usage).
7. The Final Report: A description of the software engineering processes involved in writing your game.
8. The Final Report: A description of interesting techniques, hacks or data structures you used (or could have used) in writing your game.

Suggested Extensions

• Use of AI, machine learning on the results of the simulation to improve performance.
• Use of a camera or bluetooth or other libraries.
• High Score stored securely off-phone to enable competition
• The goal here is to push the envelop of the project, expand the breadth of technologies and concepts you'd need to learn to complete the game.
• Perhaps: Using complex game AI, perhaps also using machine learning to improve the performance of that AI.
• Perhaps: Using peripheral devices for interaction with the game.
• Perhaps: Enabling multiplayer modes with networking, perhaps through scoreboards or other means.

Reading

• Mario Zechner. Beginning Android Games. 2011 - Springer.
• http://www.gameprogrammingpatterns.com/contents.html - Game Programming Patterns by Robewrt Nystrom

Early Deliverables

1. A simple proof-of-concept implementation of the aggregating algorithm and fixed share algorithm.
2. Report: An overview of the problem of prediction with expert advice and the work of the aggregating algorithm.
3. Report: The performance of the aggregating algorithm on a small artificial dataset.

Final Deliverables

1. The program will work with a real dataset, read the data from a file, preprocess it, and use a standard algorithm to generate experts' predictions (if necessary).
2. The program will implement aggregating algorithm plus possibly fixed share and the aggregating algorithm for sleeping experts. The results will be visualised.
3. The report will describe implementation issues (such as the choice of data structures, numerical methods etc) necessary to apply the theory.
4. The report will describe computational experiments, analyse their results and draw conclusions.

Suggested Extensions

• Application of sleeping experts for predicting the implied volatility of options.
• Experiments with the weak aggregating algorithm.

Reading

• Y.Kalnishkan, The AA And Laissez-Faire Investment http://onlineprediction.net/?n=Main.TheAAAndLaissez-FaireInvestment
• Kalnishkan, Y. Prediction with expert advice for a finite number of experts: A practical introduction. Pattern Recognition, 126 (2022).
• Vovk, V. Competitive on‐line statistics. International Statistical Review, 69(2), 213-248 (2001).
• M.Herbster and M.Warmuth, Tracking the Best Expert, Machine Learning, 32, 151-178 (1998)
• Y.Kalnishkan and M.V.Vyugin. The weak aggregating algorithm and weak mixability. Journal of Computer and System Sciences, 74(8): 1228--1244 (2008)

Early Deliverables

1. Study conformal prediction theory and meaning of its validity and efficiency properties.
2. Pick artificial or natural data examples with various level of difference between the data segments.
3. Implement Conformal Prediction with few versions of conformity score functions and compare their performance on these data examples.

Final Deliverables

1. Survey of conformity scores for CP framework based on standard machine learning algorithms.
2. For one of the algorithms, develop and compare different ways to define conformity score function.
3. Compare their validity and efficiency on the data examples including the transfer learning challenge.

Suggested Extensions

• To try the same for other underlying algorithms, and to formulate general principles behinds safe conformity score functions

Reading

• Hedging predictions in machine learning by Alex Gammerman and Vladimir Vovk. Computer Journal 50:151--177 (2007).
• A tutorial on conformal prediction by Glenn Shafer and Vladimir Vovk. Journal of Machine Learning Research 9:371--421 (2008).

Early Deliverables

1. A report giving an overview of prominent distributed multi-shot consensus algorithms.
2. A report describing consensus algorithms in the context of prominent permissioned and/or permissionless Blockchains.
3. A proof-of-concept program implementing a simple (possibly) inefficient multi-shot consensus algorithm

Final Deliverables

1. A permissioned Blockchain that is resilient to crash and/or Byzantine failures and achieves a high transaction throughput.
2. A report describing the architecture of the implemented Blockchain
3. A report on the performed experiments and performance evaluation results

Suggested Extensions

• Support for nodes joining and leaving the system.
• An integration with the Jepsen distributed testing framework.

Reading

• Paxos made simple, L. Lamport, ACM SIGACT News, December 2001, https://www.microsoft.com/en-us/research/publication/paxos-made-simple
• The Part-Time parliament, L. Lamport, ACM TOCS '98: https://www.microsoft.com/en-us/research/publication/part-time-parliament/
• Lecture notes on Leader-based Sequence Paxos - an understandable sequence consensus algorithm, S. Haridi et al, https://arxiv.org/pdf/2008.13456.pdf
• In search of an understandable consensus algorithm, D. Ongaro et al, USENIX ATC '14 https://www.usenix.org/conference/atc14/technical-sessions/presentation/ongaro
• Streamlet: Textbook Streamlined Blockchains, B. Chan and E. Shi, https://eprint.iacr.org/2020/088
• Practical Byzantine fault-tolerance, M. Castro and B. Liskov, OSDI '99: http://pmg.csail.mit.edu/papers/osdi99.pdf

Early Deliverables

1. Proof of concept program: Gaussian mixture model applied to a small artificial dataset.
2. A report giving an overview of clustering and describing three different clustering methods.
3. A report describing and assessing the results of applying the proof of concept program to data.

Final Deliverables

1. The program(s) will work with a real dataset(s).
2. You will generate and compare different clusterings or generative models using several different methods, such as: kernel-based semi-supervised clustering, spectral learning, factor analysis, tree-based methods, or other comparable methods.
3. You will compare and visualize the clusterings produced.
4. The report will describe and derive the formulae for the methods that you use, and the report will also discuss the implementation.

Suggested Extensions

• There is a very wide range of clustering and generative-modelling algorithms in standard use, and an even wider variety of methods of fitting them to data. This would be an opportunity to implement several different methods and get to understand them by using them in practice.

Reading

• Pattern Recognition and Machine Learning, by C.M. Bishop, Springer 2006
• Machine learning: a probabilistic perspective”, by K. Murphy, MIT Press 2012
• S. Basu, A. Bamerjee and R. Mooney, "Semi-supervised clustering by seeding", Proceedings of the 19th International Conference on Machine Learning, pp.19-26, 2002.

Early Deliverables

1. A review and a report of the approach in the case of classification and regression.
2. A design of the system and its extension for conformal predictors.
3. Implementation of a pilot system and some preliminary experiments.

Final Deliverables

1. Implementation of the system, experiments with various parameters.
2. Experimenting with different underlying ML algorithms.
3. It is desirable that the final system will have a graphical user interface.
4. The final report will describe the theory, the implementation issues, computational experiments with different data sets and parameters.

Suggested Extensions

• Online learning with conformal predictors.

Reading

• V.Vovk, A.Gammerman, G.Shafer. "Algorithmic Learning in a Random World", Springer, 2005.

Early Deliverables

1. Find and choose a data set
2. Implement a relevant version of Conformal Prediction algorithm
3. Impute some mistakes into the data, apply the algorithm, and observe the results.

Final Deliverables

1. Implement a Conformal Prediction with an additional function of error detection.
2. Compare different versions of Conformal Prediction by their efficiency for data correction.
3. Analyse and report the output.

Suggested Extensions

• Propose, justify theoretically, and implement a scheme of multiple error correction.
• Compare how the approach works in different (on-line, batch, cross-validation, inductive) modes.

Reading

• Hedging predictions in machine learning by Alex Gammerman and Vladimir Vovk. Computer Journal 50:151--177 (2007).
• A tutorial on conformal prediction by Glenn Shafer and Vladimir Vovk. Journal of Machine Learning Research 9:371--421 (2008).
• Mushroom Data Set. https://archive.ics.uci.edu/ml/datasets/Mushroom

Early Deliverables

1. Study data sets and identify one relevant for the project, describe it and the structure of its label space.
2. Select a basic machine learning approach applicable to this data.
3. Develop and implement a multi-class version of this approach and get some initial results on the data

Final Deliverables

1. Survey of the data and machine learning techniques.
2. Develop a conformity score functions for data with structured label space.
3. Implement a conformal prediction algorithm applicable to the data.
4. Analysis of the output.

Suggested Extensions

• Compare results with different basic machine learning approaches.
• Compare original label space structure attached to the data against its automatic structuring by means of data analysis.

Reading

• Hedging predictions in machine learning by Alex Gammerman and Vladimir Vovk. Computer Journal 50:151--177 (2007).
• A tutorial on conformal prediction by Glenn Shafer and Vladimir Vovk. Journal of Machine Learning Research 9:371--421 (2008).
• Predicting structured objects with support vector machines by T.Joachims, T.Hofmann, Y.Yue, C.N.Yu. Communications of the ACM 52 (11), 97--104

Early Deliverables

1. Producing a simple proof-of-concept implementation of classification-based conformal prediction with different underlying algorithms (e.g. SVM, k-NN, Random Forests, etc.) from scratch.
2. A report of the empirical results of such conformal predictors on a small synthetic dataset.

Final Deliverables

1. A detailed literature review of existing machine learning algorithms used for non-speech sound recognition.
2. Producing a detailed report of the dataset structure, including some visual analysis (e.g. t-SNE, UMAP, etc.).
3. Implementing conformal predictors on the whole dataset, optimising the underlying algorithms' parameters.
4. Carefully analysing the p-values, comparing the validity, efficiency, accuracy of the implemented conformal predictors.

Suggested Extensions

• Comparing different methods of extracting sound features (e.g. MFCC, CFA, ZCR) from raw recordings, and their impacts on the prediction performance.
• Using the entire dataset to train a new model, then testing such model on other real world recordings.

Reading

• Nguyen, K. A. and Luo, Z. (2018). Cover Your Cough: Detection of Respiratory Events with Confidence Using a Smartwatch. Proceedings of Machine Learning Research, 91, 1-18.
• Piczak, K. J. (2015). ESC: Dataset for environmental sound classification. In Proceedings of the 23rd ACM international conference on Multimedia (pp. 1015-1018). ACM.
• https://dataverse.harvard.edu/dataset.xhtml?persistentId=doi:10.7910/DVN/YDEPUT

Early Deliverables

1. A report on Hidden Markov Model (HMM) tools in Bioinformatics.
2. A report on conformal predictors.
3. A simple implementation of conformal prediction on a family of proteins from PFAM.

Final Deliverables

1. A framework to evaluate the credibility and confidence of individual HMM's from PFAM.

Suggested Extensions

• Using conformal predictors on HMM models in protein sequence prediction.

Reading

• The Pfam protein families database. Finn RD1, Mistry J, Tate J, Coggill P, Heger A, Pollington JE, Gavin OL, Gunasekaran P, Ceric G, Forslund K, Holm L, Sonnhammer EL, Eddy SR, Bateman A. doi: 10.1093/nar/gkp985.
• http://jmlr.csail.mit.edu/papers/volume9/shafer08a/shafer08a.pdf

Early Deliverables

1. A proof-of-concept program implementing neural nets both in the case of regression and in the case of classification (where neural networks output probability predictions).
2. Building a cross-conformal predictor on top of neural nets and exploring its empirical validity.
3. A preliminary empirical study of the efficiency of the student's implementation and standard implementations (such as the one in scikit-learn).

Final Deliverables

1. Careful cross-validation of various parameters, including the stopping rule, learning rate, and the numbers of hidden neurons at various levels.
2. Implementing regularization.
3. Experimenting with different architectures, possibly including deep architectures.
4. Ideally, the final product will have a graphical user interface.
5. The final report will describe the theory behind the algorithms, the implementation issues necessary to apply the theory, the software engineering process involved in generating your software, computational experiments with different data sets, methods, and parameters.

Suggested Extensions

• Careful study of the optimal number of folds.
• Replacing back-propagation with other optimization algorithms.
• Implementing jackknife+, a popular recent version of cross-conformal prediction.

Reading

• Hastie, Tibshirani, and Friedman, The Elements of Statistical Learning, 2009, Chapter 11.
• Mitchell, Machine Learning, 1997, Chapter 4.
• Research papers.

Early Deliverables

1. Proof of concept program: Implementation of a deep learning algorithm for a relatively small network on a relatively simple dataset, together with experiments determining the effect of hyperparameters on convergence
2. A report giving an overview of a range of related deep learning methods.
3. A report describing and assessing the performance of the initial implementation and any experiments performed

Final Deliverables

1. The program(s) will work with a real dataset(s).
2. You will implement and compare multiple architectures and optimisation methods
3. You will visualize and assess the performance of the algorithms produced.
4. The report will describe and derive the formulae for the methods that you use, and the report will also discuss the implementation.

Suggested Extensions

• This would be an opportunity to implement several deep learning methods and get to understand them in practice by applying them to a significant dataset.

Reading

• "Deep Learning", by Goodfellow, Bengio, and Courville, MIT Press 2016
• "Deep Learning with Python" by Francois Chollet, Manning Publications 2018

Early Deliverables

1. A proof-of-concept implementation of bagging, random forests, or boosting both in the case of regression and in the case of classification.
2. Implementing underlying algorithms, such as decision stumps or decision trees; simple proof-of-concept implementations are sufficient.

Final Deliverables

1. A final program implementing bagging, random forests, or boosting (or all three) with a careful choice of the parameters.
2. Implementations of the underlying algorithms should be complemented by a careful empirical study of the effect of various parameters; producing optimized programs.
3. Comparing the performance of different ensemble techniques in the problems of classification and regression.
4. Using bagged and boosted decision trees as probability predictors; evaluating their perfomance using proper loss functions.
5. Ideally, the final product will have a graphical user interface.
6. The final report will describe the theory behind the algorithms, the implementation issues necessary to apply the theory, the software engineering process involved in generating your software, computational experiments with different data sets, methods, and parameters.

Suggested Extensions

• Implementing the out-of-bag estimate of error probability.
• Investigating variable importance.
• Proximity plots.
• Theoretical analysis of the algorithms.

Reading

• Hastie, Tibshirani, and Friedman, The Elements of Statistical Learning, 2009, Section 8.7, Chapters 10 and 15-16
• James, Witten, Hastie, and Tibshirani, An Introduction to Statistical Learning, 2013, Chapter 8.
• Research papers.

Early Deliverables

1. A report on graphs and scale-free networks.
2. A report on the PID graph.
3. A report on GraphQL.

Final Deliverables

1. The development of a tool to query the PID graph and visualise graphs.
2. Identifying the behaviour of the degree of the PID graph.

Suggested Extensions

• Exploring other graph obseravbles e.g. centrality.
• Exploring using a recommender system to identify papers of interest.

Reading

• Introducing the PID Graph - https://blog.datacite.org/introducing-the-pid-graph/
• The Fullstack Introduction for GraphQL - https://www.howtographql.com/
• Learning GraphQL, E.Porcello and A. Banks O'Reilly Media (available online through Safari Books)
• Handbook of Graph Theory, J.L. Gross, J. Yellen and P. Zhang, Chapman and Hall/C (available online through Safari Books)
• Scale-free networks are rare, A.D. Broido and A. Clauset, Nature Comms. (2019) DOI - 10.1038/s41467-019-08746-5

Early Deliverables

1. Proof of concept program: identify one machine learning task suitable for characters recognition and implement one learning algorithm for the chosen task.
2. Report: An overview of the datasets. Statistics and numbers involved.
3. Report: A report on machine learning tasks for character recognition.

Final Deliverables

1. Survey of machine learning techniques used for characters recognition.
2. The developed machine learning software for Greek characters recognition.
3. The software engineering process involved in generating your software.

Suggested Extensions

• Provide predictions with confidence measure.

Reading

• Gatos, Basilis, et al. GRPOLY-DB: An old Greek polytonic document image database. Document Analysis and Recognition (ICDAR), 2015 13th International Conference on. IEEE, 2015.
• http://www.iit.demokritos.gr/~nstam/GRPOLY-DB

Early Deliverables

1. Proof of concept program pre-processing the data and applying two machine learning methods to Boston Housing dataset
2. Report: A report describing supervised learning, the concepts 'training set' and 'test set', and the theory of the methods.

Final Deliverables

1. A machine learning method (such as Ridge Regression; the choice is by discussion with the supervisor) should be implemented from scratch.
2. Several different datasets should be used
3. Tests comparing different kernels, parameters, etc should be performed. The results should be visualised.
4. Prediction in the online mode should be implemented and results visuaised.
5. The report will describe different algorithms and implementation issues (such as the choice of data structures, numerical methods etc) necessary to apply the theory.
6. The report will describe computational experiments, analyse their results and draw conclusions

Suggested Extensions

• Methods of prediction with expert advice can be applied in the online mode.
• Is there inflation? A study can be undertaken.

Reading

• Boston housing dataset on Kaggle: https://www.kaggle.com/prasadperera/the-boston-housing-dataset
• California housing dataset on Kaggle: https://www.kaggle.com/camnugent/california-housing-prices
• De Cock, Dean. Ames, Iowa: Alternative to the Boston housing data as an end of semester regression project. Journal of Statistics Education 19.3 (2011).
• Y.Kalnishkan, Kernel Methods, http://onlineprediction.net/?n=Main.KernelMethods
• J.Shawe-Taylor and N.Cristianini, Kernel Methods for Pattern Analysis Cambridge University Press, 2004
• B.Schlkopf and A.J.Smola, Learning with Kernels: Support Vector Machines, Regularization, Optimization, and Beyond, The MIT Press, 2001.

Early Deliverables

1. A report on Gene Ontologies.
2. A report on the tool GOCat explaining the methods.
3. A report on the data set required to build the tool.

Final Deliverables

1. The development of a stand alone tool to implement the equivalent of GOCat.
2. An API to query the tool.
3. Rigourous testing of the tool.

Suggested Extensions

• Using more advanced methods beyond kNN for the implementation.
• An API to query the tool.

Reading

• Managing the data deluge: data-driven GO category assignment improves while complexity of functional annotation increases. Gobeill J Pasche E Vishnyakova D Ruch P DOI 10.1093/database/bat041
• A survey on annotation tools for the biomedical literature. Neves M Leser U DOI 10.1093/bib/bbs084
• The GOA database in 2009--an integrated Gene Ontology Annotation resource Barrell D Dimmer E Huntley R Binns D O'Donovan C Apweiler R dx.doi.org/10.1093/nar/gkn803

Early Deliverables

1. Implementation of transfer learning algorithms based on pre-trained networks for any of the following tasks, (1) object detection/classification, (2) image segmentation, and (3) video classification.
2. A report describing transfer learning algorithms implemented with some evaluation results.

Final Deliverables

1. Implementing transfer learning, deep learning (e.g. Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs)), or hybrid deep networks for any of the aforementioned tasks.
2. Implementing several deep learning algorithms with different architectures, where the employed architectures are either existing, or newly proposed.
3. Implementing manual hyper-parameter fine-tuning of the implemented deep networks.
4. A comprehensive evaluation and comparison against existing state-of-the-art studies.
5. The final report will describe methodology, implementation and testing with theoretical analysis and detailed evaluation results.

Suggested Extensions

• Implementing deep hybrid networks with new architectures.
• Implementing optimal hyper-parameter selection using an automatic process.
• Theoretical analysis of the algorithms.

Reading

• Goodfellow, I., Bengio, Y. and Courville, A., 2016. Deep learning. MIT press.
• Chollet, F., 2021. Deep learning with Python.
• Journal/conference papers.

Early Deliverables

1. Report 1: bootstrap sampling.
2. Report 2: overview of chosen case study.
3. Report 3: prediction regions, JPRs and bootstrap JPRs.
4. Report 4: time-series models.
5. Report 5: preliminary plan, including an overview of the planned solution method and experimental evaluation plan.
6. Implement and evaluate predictor and prediction standard errors.

Final Deliverables

1. Implement code for bootstrap JPRs and evaluate empirical coverage and family-wise error rate for different hyper-parameters (e.g., k, horizon length, sample size, number of bootstrap samples).
2. Final dissertation describing problem, methods and algorithm, implementation details, computational experiments and their analysis, and conclusions.

Suggested Extensions

• 1. Compare different predictors.
• 2. Compare bootstrap JPRs with other methods mentioned in the paper (e.g., joint marginals and Scheffé)
• 3. Compare bootstrap JPRs with Monte Carlo sampling based on Bayesian neural networks and uncertainty recalibration

Reading

• (bootstrap) Chapter 6, Lecture notes of 36-402, CMU Undergraduate Advanced Data Analysis, available at https://www.stat.cmu.edu/~cshalizi/ADAfaEPoV/ADAfaEPoV.pdf
• (bootstrap JPRs) Wolf, Michael, and Dan Wunderli. "Bootstrap joint prediction regions." Journal of Time Series Analysis 36.3 (2015): 352-376.
• (time-series models) Stock Market Predictions with LSTM in Python, available at https://www.datacamp.com/community/tutorials/lstm-python-stock-market
• (time-series models) Zhang, A., Lipton, Z. C., Li, M., & Smola, A. J. (2019). Dive into deep learning, chapters 8 and 9, available at https://d2l.ai/index.html
• (extension 3) Gal Y. & Ghahramani Z., 2016, Dropout as a Bayesian Approximation: Representing Model Uncertainty in Deep Learning, ICML 2016, available at http://proceedings.mlr.press/v48/gal16.pdf
• (extension 3) Kuleshov, Volodymyr, Nathan Fenner, and Stefano Ermon. "Accurate Uncertainties for Deep Learning Using Calibrated Regression." ICML 2018, available at https://arxiv.org/pdf/1807.00263.pdf

Early Deliverables

1. Report on the comparative study of the distributed platforms for processing large-scale graphs.
2. Implementation and evaluation of the proof-of-concept programs.
3. Detailed plan for the final project outcomes as agreed with the advisor.

Final Deliverables

1. The software you built to achieve the agreed project goals. The dissertation will report on different algorithms for the advanced problem you chose comparing their strengths and weaknesses. It will discuss techniques and findings, and provide meaningful visualisation of the results.

Reading

Early Deliverables

1. Report 1: overview of the artificial pancreas and underlying control algorithms.
2. Report: review of main models for glucose prediction.
3. Report: preliminary plan, including an overview of the planned solution method and experimental evaluation plan.
4. Implement code for data processing and for learning neural network-based glucose predictors. Perform training for at least one NN architecture and at least one prediction horizon.

Final Deliverables

1. Final dissertation describing problem, data, solution method, machine learning methods used, implementation details, computational experiments and their analysis, and conclusions.

Suggested Extensions

• Analysis of prediction reliability using techniques like inductive conformal prediction or approximate Bayesian inference via Monte Carlo dropout.
• Derive optimal insulin control policies using the learned glucose prediction models
• Compare the learned deep models with more classical autoregressive models.

Reading

• (for report 1) Boughton, Charlotte K., and Roman Hovorka. "Advances in artificial pancreas systems." Science translational medicine 11.484 (2019). PDF available from supervisor
• (for reports 1 and 2) Lunze, Katrin, et al. "Blood glucose control algorithms for type 1 diabetic patients: A methodological review." Biomedical signal processing and control 8.2 (2013): 107-119.
• (for report 2 and for ideas on neural net architecture) Li, Kezhi, et al. "Convolutional recurrent neural networks for glucose prediction." IEEE Journal of Biomedical and Health Informatics (2019).
• (for report 2 and for example of autoregressive model) Sparacino, Giovanni, et al. "Glucose concentration can be predicted ahead in time from continuous glucose monitoring sensor time-series." IEEE Transactions on biomedical engineering 54.5 (2007): 931-937.
• (for monte carlo dropout) Gal, Yarin, and Zoubin Ghahramani. "Dropout as a bayesian approximation: Representing model uncertainty in deep learning." international conference on machine learning. 2016.
• (for inductive conformal prediction) Papadopoulos, Harris. "Inductive conformal prediction: Theory and application to neural networks." Tools in artificial intelligence. IntechOpen, 2008.

Early Deliverables

1. Proof of concept program: identify one machine learning task suitable for the smart meter data and implement one learning algorithm for the chosen task.
2. Report: An overview of the datasets used. Statistics and numbers involved.
3. Report: A report on machine learning tasks for smart meter data.

Final Deliverables

1. Survey of machine learning techniques used for smart meter data analysis.
2. The developed machine learning software for smart meter data.
3. The software engineering process involved in generating your software.
4. Experiments with different smart meter data.

Suggested Extensions

• Extend the learning algorithms for online setting.
• Provide predictions with confidence measure.

Reading

• Smart*: An Open Data Set and Tools for Enabling Research in Sustainable Homes. Sean Barker, Aditya Mishra, David Irwin, Emmanuel Cecchet, Prashant Shenoy, and Jeannie Albrecht. Proceedings of the 2012 Workshop on Data Mining Applications in Sustainability (SustKDD 2012), Beijing, China, August 2012.
• Smart* Data Set for Sustainability http://traces.cs.umass.edu/index.php/Smart/Smart
• SmartMeter Energy Consumption Data in London Households https://data.london.gov.uk/dataset/smartmeter-energy-use-data-in-london-households
• UK Domestic Appliance-Level Electricity (UK-DALE) dataset http://jack-kelly.com/data/
• Revealing Household Characteristics from Smart Meter Data. Christian Beckel, Leyna Sadamori, Thorsten Staake, Silvia Santini, Energy, Vol. 78, pp. 397-410, December 2014.
• Piotr Mirowski, Sining Chen, Tin Kam Ho, Chun-Nam Yu, Demand Forecasting in Smart Grids, Bell Labs Technical Journal, 18(4), pp.135-158, 2014.

Early Deliverables

1. Producing a simple proof-of-concept implementation of some classification and regression algorithms from scratch.
2. A report of the empirical results of such learning algorithms on a small synthetic dataset.

Final Deliverables

1. A detailed literature review of the machine learning application in WiFi fingerprinting.
2. Producing a detailed report of the dataset structure, including visual analysis (e.g. t-SNE, UMAP, etc.).
3. Implementing the algorithms on the entire competition dataset, and optimising their parameters.
4. Carefully analysing and comparing the performance accuracy of such learning algorithms.

Suggested Extensions

• Providing location estimations with some confidence measure.
• Extending the learning algorithms to an online setting, where the user provides continuous samples as she traverses the building.

Reading

• Bahl, P. and Padmanabhan, V. N. (2000). RADAR: An in-building RF-based user location and tracking system. In INFOCOM 2000. 19th Annual Joint Conference of the IEEE Computer and Communications Societies. Proceedings. IEEE (Vol. 2, pp. 775-784).
• Honkavirta, V., Perala, T., Ali-Loytty, S., and Piché, R. (2009). A comparative survey of WLAN location fingerprinting methods. In Positioning, Navigation and Communication, 2009. WPNC 2009. 6th Workshop on (pp. 243-251). IEEE.
• https://archive.ics.uci.edu/ml/datasets/UJIIndoorLoc

Early Deliverables

1. Proof of concept program: tf-idf based similarity measurement
2. Proof of concept program: word-embedding based similarity measurement
3. (Optional) Proof of concept program: sentence-embedding based similarity measurement.
4. A report describing and assessing the results of applying the proof of concept programmes on different sentence pairs.

Final Deliverables

1. Fine-tune existing neural models (e.g. BERT), and evaluate whether they are better at dealing problems (i) and (ii) than other models.
2. A report summarising all the experiments and analysing strengths and weaknesses of each model

Suggested Extensions

• Constructing more training sentence pairs, for example by using Google Translate: given a sentence in English (denoted S), first translate it into another language (denote the translated sentence by S'), and then translate S' back to English (denote the result as S''); we should assume that S' and S'' should be similar sentences.

Reading

• Neural Network Methods for Natural Language Processing, by Y. Goldberg, Morgan and Claypool 2017
• Deep Learning, by I. Goodfellow, Y. Bengio and A. Courville, MIT Press 2016
• J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova, "BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding", arXiv pre-print 1810.04805

Early Deliverables

1. Proof of concept program: Long Short Term Memory (LSTM) Neural Network for text encoding.
2. (Optional) Proof of concept program: Attention Mechanism on top of (LSTM).
3. A report describing and assessing the results of applying the proof of concept programmes on data.

Final Deliverables

1. Implement an LSTM-based Encoder-Decoder network and train it with a few hundred thousand of real data
2. Augment the Encoder-Decoder network with Pointer Mechanism and Attention Mechanism, compare their performance against vanilla Encoder-Decoder
3. (Optional) Implement Transformer-based Encoder-Decoder, compare its performance against the other models
4. A report summarising all the experiments and analysing strengths and weaknesses of each model

Suggested Extensions

• Applying Reinforcement Learning (RL) on top of the Encoder-Decoder models. Recent research suggests that applying RL further boost the performance.
• Using Transfer Learning techniques to adjust existing language models, e.g. GPT-2, to generate summaries.

Reading

• Neural Network Methods for Natural Language Processing, by Y. Goldberg, Morgan and Claypool 2017
• Deep Learning, by I. Goodfellow, Y. Bengio and A. Courville, MIT Press 2016
• A. See, P. J. Liu and C. Manning, "Get To The Point: Summarization with Pointer-Generator Networks", Proceedings of the 55th ACL Annual Conference

Early Deliverables

1. Precise specification of the project goals as agreed with the advisor along with data collection and analysis tools, and methodologies you plan to use to achieve the specified goals.
2. Intermediate report and programs showing and reflecting upon initial findings.

Final Deliverables

1. The software you built to achieve the project goals.
2. Final report will discuss techniques and findings, and provide meaningful visualisation of the results.

Reading

• Brendan O’Connor, Ramnath Balasubramanyan, Bryan R. Routledge, Noah A. Smith, "From Tweets to Polls: Linking Text Sentiment to Public Opinion Time Series" In Proceedings of the International AAAI Conference on Weblogs and Social Media (2010)

Early Deliverables

1. Proof of concept program: Page rank applied to a small (possibly synthetic) dataset. Its performance and efficiency should be analysed.
2. Report: An overview of the problem of document ranking and main approaches to it. An overview of PageRank.

Final Deliverables

1. Two or more different machine learning technique should be applied to a standard benchmark dataset and their performance compared.
2. The performance of the program should be optimised to enable them to process large amounts of data.
3. The report will describe the theory of underlying learning algorithms.
4. The report will describe implementation issues (such as the choice of data structures, numerical methods etc) necessary to apply the theory.
5. The report will describe computational experiments, analyse their esults and draw conclusions.

Suggested Extensions

• Comparison of pointwise, pairwise and listwise approaches.
• Use of real Internet dumps such as http://commoncrawl.org/

Reading

• Christiane Rousseau, How Google works: Markov chains and eigenvalues, URL http://blog.kleinproject.org/?p=280
• Chuan He, Cong Wang, Yi-xin Zhong, and Rui-Fan Li. A survey on learning to rank. Machine Learning and Cybernetics, 2008 International Conference on. DOI 10.1109/ICMLC.2008.4620685.
• Tie-Yan Liu, Learning to Rank for Information Retrieval, Berlin, Heidelberg: Springer. 2011

Early Deliverables

1. Proof of concept program: Kernel Ridge Regression applied to a small artificial dataset.
2. Report: An overview of ridge regression describing the concepts 'training set' and 'test set', giving the formulas for Ridge Regression and defining all terms in them.
3. Report: Examples of applications of Ridge Regression worked out on paper and checked using the prototype program.

Final Deliverables

1. The program will work with a real dataset such as Boston housing, read the data from a file, preprocess the data (normalise, split the dataset into the test and training sets) and apply Ridge Regression using a kernel with parameters.
2. The program will automatically perform tests such as comparison of different kernels, parameters, etc. The results will be visualised.
3. The program will work in batch and on-line modes.
4. Tests will be performed to compare plain Ridge Regression against other regression-based algorithms such as Kernel Aggregating Algorithm Regression, Kernel Aggregating Algorithm Regression with Changing Dependencies, or gradient descent.
5. The report will describe the theory of Ridge Regression and derive the formulas.
6. The report will describe implementation issues (such as the choice of data structures, numerical methods etc) necessary to apply the theory.
7. The report will describe computational experiments, analyse their results and draw conclusions

Suggested Extensions

• Application of Ridge Regression to different datasets (including those suggested by the student).
• Comparison with other learning algorithms (e.g., aggregating algorithm regression, neural networks, gradient descent-based etc).
• Combining regression with methods of prediction with expert advice.

Reading

• Y.Kalnishkan, Kernel Methods, http://onlineprediction.net/?n=Main.KernelMethods
• J.Shawe-Taylor and N.Cristianini, Kernel Methods for Pattern Analysis Cambridge University Press, 2004
• C.E.Rasmussen and C.Williams, Gaussian Processes for Machine Learning, the MIT Press, 2006 http://www.gaussianprocess.org/
• B.Schlkopf and A.J.Smola, Learning with Kernels: Support Vector Machines, Regularization, Optimization, and Beyond, The MIT Press, 2001.

Early Deliverables

1. Report: overview of SAT, SMT and SMT solving techniques.
2. Report: bounded model checking and encoding into SMT.
3. Install z3 and Python APIs. Develop code for SMT solving of a simple example problem (e.g., Sudoku, or job shop scheduling).
4. Develop code for BMC of generic transition systems.

Final Deliverables

1. Develop code for encoding a planning problem on a 2-d grid with obstacle as a BMC problem.
2. Evaluate your solution with arbitrary and randomly generated planning instances.
3. Analyze SMT solver performance for different grid sizes and numbers of obstacles.
4. Final dissertation. It should include background on SAT, SMT, BMC, and AI planning, as well as the description of your solution for the encoding of AI planning into SMT, some example solutions obtained with the developed planning method, and performance analysis results.

Suggested Extensions

• Extend the planning problem and your solution with transition costs and cost constraints.
• Develop code to visualize path solutions on the 2d grid.

Reading

• De Moura, Leonardo, and Nikolaj Bjørner. "Satisfiability modulo theories: introduction and applications." Communications of the ACM 54.9 (2011): 69-77. (available at https://dl.acm.org/doi/abs/10.1145/1995376.1995394)
• De Moura, Leonardo, and Nikolaj Bjørner. "Satisfiability modulo theories: An appetizer." Brazilian Symposium on Formal Methods (2009): 23-36. (available at https://link.springer.com/chapter/10.1007/978-3-642-10452-7_3)
• Amla, Nina, et al. "An analysis of SAT-based model checking techniques in an industrial environment." (up to section 2.3 included). (available at https://link.springer.com/chapter/10.1007/11560548_20)
• Lectures on " Introduction to planning" and " Planning formalisms" of the " Autonomous Intelligent Systems" course
• Z3 tool page, https://github.com/Z3Prover/z3
• Programming Z3 tutorial (sections 1,2,3,5), https://theory.stanford.edu/~nikolaj/programmingz3.html
• Jupyter notebooks illustrating Z3's Python APIs (includes code for Programming Z3 tutorial), https://notebooks.azure.com/nbjorner/projects/z3examples
• Online Z3 tutorial (using SMT-lib input format), https://rise4fun.com/z3/tutorial/guide
• Dennis Yurichev, "SAT/SMT by Examples", https://yurichev.com/writings/SAT_SMT_by_example.pdf (a huge collection of maths and CS problems solved with SAT or SMT)

Early Deliverables

1. A report giving an overview of prominent existing key value stores such as Chord and SkipGraph.
2. A report describing how existing key value stores deal with environmental changes.
3. A proof-of-concept program implementing a simple (possibly inefficient) distributed key value store.

Final Deliverables

1. A distributed key value store that provides standard dictionary operations and is scalable to large data sets.
2. A report describing the architecture of the implemented key value store.
3. A report on the performed experiments and simulation results.

Suggested Extensions

• Supporting advanced dictionary operations, e.g., range queries.
• More efficient storage of data by using coding techniques.
• Simulating node churn.

Reading

• Ion Stoica, Robert Morris, David R. Karger, M. Frans Kaashoek, Hari Balakrishnan: Chord: A scalable peer-to-peer lookup service for internet applications. SIGCOMM 2001: 149-160
• James Aspnes, Gauri Shah: Skip graphs. ACM Trans. Algorithms 3(4): 37 (2007)

Early Deliverables

1. You will write a simple proof-of-concept program for computing Value at Risk for a portfolio consisting of 1 or 2 stocks using two different methods: model-building and historical simulation.
2. Different ways of computing Value at Risk will be back tested and the statistical significance of the results analyzed.
3. The report will describe the program (software engineering, algorithms, etc.); discuss different ways of estimating the variance of returns and the covariance between returns of two different securities (using the standard formula, exponentially weighted moving average, or GARCH(1,1)).

Final Deliverables

1. The program should be extended by allowing derivatives (such as options) in the portfolio and using Monte Carlo simulation. Option pricing can be performed using the Black-Scholes formula, or binomial trees (for American options), or a different variety of Monte Carlo simulation (for Asian options).
2. Allow an arbitrary number of stocks in the model-building approach (using eigen- or Cholesky decomposition).
3. Allow an arbitrary number of stocks using histotical simulation.
4. Ideally, it will have a graphical user interface.
5. The final report will describe the theory behind the algorithms; the implementation issues necessary to apply the theory; the software engineering process involved in generating your software; computational experiments with different data sets, methods, and parameters.

Suggested Extensions

• Computing conditional VaR (also known as expected shortfall)
• Explore carefully the choice of parameters for EWMA and GARCH(1,1) using a loss function such as square or log loss
• Empirical study of the two approaches to historical simulation for n-day VaR: reducing to 1-day VaR and direct estimation
• Complement back testing by stress testing.
• Replacing the Gaussian model for stock returns by robust models.
• Computing monetary measures of risk different from VaR.
• Using proper scoring rules for quantiles for valuating computed VaRs.
• Using methods of worst-case risk aggregation, such as the Rearrangement Algorithm, or worst-case online decision making, such as the Weak Aggregating Algorithm (currently active areas of research).

Reading

• John C. Hull. Options, futures and other derivatives, 10th ed., Pearson, Upper Saddle River, NJ, 2018.
• Hans Follmer and Alexander Schied. Stochastic Finance: An Introduction in Discrete Time 3rd ed., de Gruyter, Berlin, 2011.
• Yahoo Finance as source of data: http://finance.yahoo.com/
• http://gloria-mundi.com

Early Deliverables

1. Proof of concept program: Gaussian mixture model applied to a small artificial dataset.
2. A report giving an overview of clustering and describing three different clustering methods.
3. A report describing and assessing the results of applying the proof of concept program to data.

Final Deliverables

1. The program(s) will work with a real dataset(s).
2. You will generate and compare different clusterings or generative models using several different methods, such as: kernel-based semi-supervised clustering, spectral learning, factor analysis, tree-based methods, or other comparable methods.
3. You will compare and visualize the clusterings produced.
4. The report will describe and derive the formulae for the methods that you use, and the report will also discuss the implementation.

Suggested Extensions

• There is a very wide range of clustering and generative-modelling algorithms in standard use, and an even wider variety of methods of fitting them to data. This would be an opportunity to implement several different methods and get to understand them by using them in practice.

Reading

• Pattern Recognition and Machine Learning, by C.M. Bishop, Springer 2006
• Machine learning: a probabilistic perspective”, by K. Murphy, MIT Press 2012
• S. Basu, A. Bamerjee and R. Mooney, "Semi-supervised clustering by seeding", Proceedings of the 19th International Conference on Machine Learning, pp.19-26, 2002.

Early Deliverables

1. Proof of concept program: Gaussian mixture model applied to a small artificial dataset.
2. A report giving an overview of clustering and describing three different clustering methods.
3. A report describing and assessing the results of applying the proof of concept program to data.

Final Deliverables

1. The program(s) will work with a real dataset(s).
2. You will generate and compare different clusterings or generative models using several different methods, such as: kernel-based semi-supervised clustering, spectral learning, factor analysis, tree-based methods, or other comparable methods.
3. You will compare and visualize the clusterings produced.
4. The report will describe and derive the formulae for the methods that you use, and the report will also discuss the implementation.

Suggested Extensions

• There is a very wide range of clustering and generative-modelling algorithms in standard use, and an even wider variety of methods of fitting them to data. This would be an opportunity to implement several different methods and get to understand them by using them in practice.

Reading

• Pattern Recognition and Machine Learning, by C.M. Bishop, Springer 2006
• Machine learning: a probabilistic perspective”, by K. Murphy, MIT Press 2012
• S. Basu, A. Bamerjee and R. Mooney, "Semi-supervised clustering by seeding", Proceedings of the 19th International Conference on Machine Learning, pp.19-26, 2002.

Early Deliverables

1. You will implement simple machine learning algorithms such as nearest neighbours and decision trees.
2. They will be tested using simple artificial data sets.
3. Report: a description of 1-nearest neighbour and k-nearest neighbours algorithms, with different strategies for breaking the ties;
4. Report: a description of decision trees using different measures of uniformity.

Final Deliverables

1. May include nearest neighbours using kernels and multi-class support vector machines.
2. The algorithms will be studied empirically on benchmark data sets such as those available from the UCI data repository and the Delve repository.
3. For many of these data set judicious preprocessing (including normalisation of attributes or examples) will be essential.
4. The performance of all algorithms will be explored using a hold-out test set and cross-validation.
5. The overall program will have a full object-oriented design, with a full implementation life cycle using modern software engineering principles.
6. Ideally, it will have a graphical user interface.
7. The report will describe: the theory behind the algorithms.
8. The report will describe: the implementation issues necessary to apply the theory.
9. The report will describe: the software engineering process involved in generating your software.
10. The report will describe: computational experiments with different data sets and parameters.

Suggested Extensions

• Modifications of known algorithms.
• Dealing with machine learning problems with asymmetrical errors (such as spam detection) and ROC analysis.
• Nontrivial adaptations of known algorithms to applications in a specific area, such as medical diagnosis or option pricing.
• Exploring the cross-validation procedure, in particular the leave-one out procedure. What is the optional number of folds?
• Comparative study of different strategies of reducing multi-class classifiers to binary classifiers (such as one-against-one, one-against-the-rest, coding-based).

Reading

• Trevor Hastie, Robert Tibshirani, and Jerome Friedman. The Elements of Statistical Learning. Second edition. Springer, New York, 2009.
• Tom M. Mitchell. Machine Learning. McGrow-Hill, New York, 1997.
• Vladimir N. Vapnik. The Nature of Statistical Learning Theory. Second edition. Springer, New York, 2000.
• Vladimir N. Vapnik. Statistical Learning Theory. Wiley, New York, 1998.
• Seth Hettich and Steven D. Bay. The UCI KDD Archive. University of California, Department of Information and Computer Science, Irvine, CA, 1999, http://kdd.ics.uci.edu.
• Delve archive. http://www.cs.toronto.edu/~delve.

Early Deliverables

1. You will implement simple machine learning algorithms such as nearest neighbours and decision trees.
2. They will be tested using simple artificial data sets.
3. Report: a description of 1-nearest neighbour and k-nearest neighbours algorithms, with different strategies for breaking the ties;
4. Report: a description of decision trees using different measures of uniformity.

Final Deliverables

1. May include nearest neighbours using kernels and multi-class support vector machines.
2. The algorithms will be studied empirically on benchmark data sets such as those available from the UCI data repository and the Delve repository.
3. For many of these data set judicious preprocessing (including normalisation of attributes or examples) will be essential.
4. The performance of all algorithms will be explored using a hold-out test set and cross-validation.
5. The overall program will have a full object-oriented design, with a full implementation life cycle using modern software engineering principles.
6. Ideally, it will have a graphical user interface.
7. The report will describe: the theory behind the algorithms.
8. The report will describe: the implementation issues necessary to apply the theory.
9. The report will describe: the software engineering process involved in generating your software.
10. The report will describe: computational experiments with different data sets and parameters.

Suggested Extensions

• Modifications of known algorithms.
• Dealing with machine learning problems with asymmetrical errors (such as spam detection) and ROC analysis.
• Nontrivial adaptations of known algorithms to applications in a specific area, such as medical diagnosis or option pricing.
• Exploring the cross-validation procedure, in particular the leave-one out procedure. What is the optional number of folds?
• Comparative study of different strategies of reducing multi-class classifiers to binary classifiers (such as one-against-one, one-against-the-rest, coding-based).

Reading

• Trevor Hastie, Robert Tibshirani, and Jerome Friedman. The Elements of Statistical Learning. Second edition. Springer, New York, 2009.
• Tom M. Mitchell. Machine Learning. McGrow-Hill, New York, 1997.
• Vladimir N. Vapnik. The Nature of Statistical Learning Theory. Second edition. Springer, New York, 2000.
• Vladimir N. Vapnik. Statistical Learning Theory. Wiley, New York, 1998.
• Seth Hettich and Steven D. Bay. The UCI KDD Archive. University of California, Department of Information and Computer Science, Irvine, CA, 1999, http://kdd.ics.uci.edu.
• Delve archive. http://www.cs.toronto.edu/~delve.

Early Deliverables

1. Report on classic lossless algorithms: Implement: run-length encoding; move-to-front encoding; Huffman encoding.
2. Program: GUI built with buttons etc.
3. Program: Efficient Suffix Tree implementation
4. Program: Naive implementation of Burrows-Wheeler
5. Report: How does Burrows-Wheeler work with an emphasis on writing and understanding a proof that the reverse transform works.
6. Report: Shannon information theory and lossy v lossless compression; complexity theory; pseudo-code of algorithms implemented.
7. Reports: Complexity, NP hardness and the big O notation
8. Report: The preprocessing phase for compression.
9. Program: Implement an efficient Burrows-Wheeler Transform using Suffix Trees.
10. Program: Apply the Lyndon factorisation to an input string, and apply the Burrows-Wheeler transform to each of the Lyndon factors.

Final Deliverables

1. The program will have a Graphical User Interface with which to input text files and display compression ratios for compressed files.
2. The Lyndon factorisation and Burrows-Wheeler transformation of a string will be displayed.
3. The program must have a full object-oriented design, with a full implementation life cycle using modern software engineering principles.
4. The report will describe the software engineering process involved in generating your software, interesting programming techniques and data structures used (or able to be used) on the project.
5. The report will describe the complexity analysis of your computations.
6. The report will include some results of benchmarking compression algorithms with a discussion of their meanings.
7. The report will include discussion of applications of the Burrows-Wheeler transform and Lyndon factorisation.
8. The report will include a User Manual.

Suggested Extensions

• Build a mobile Burrow-Wheeler application: this will require use of Worker Threads.
• Show the output using an effective JTree with callbacks to enable user interaction.
• Animate the Burrows-Wheeler in order to help explain its working to a student unfamiliar with the algorithm.
• Adding Digital Signatures to compressed files.

Reading

• D. Adjeroh, T. Bell, A. Mukherjee, The Burrows-Wheeler transform: data compression, suffix arrays, and pattern matching, Springer (2008) 352 pp.
• M. Burrows and D. Wheeler, A block sorting lossless data compression algorithm, Technical Report 124, Digital Equipment Corporation (1994).
• M. Salson, T. Lecroq, M. Leonard and L. Mouchard, A four-stage algorithm for updating a Burrows-Wheeler transform, Theoret. Comput. Sci. 410 (43) (2009) 4350-4359 doi:10.1016/j.tcs.2009.07.016.
• http://en.wikipedia.org/wiki/Burrows-Wheeler transform
• M. Lothaire, Combinatorics onWords, Addison-Wesley, Reading, MA, 1983; 2nd Edition, Cambridge University Press, Cambridge, 1997.
• J. P. Duval, Factorising words over an ordered alphabet, J. Algorithms 4 (1983) 363-381.
• http://en.wikipedia.org/wiki/Data_compression

Early Deliverables

1. Report on classic lossless algorithms: Implement: run-length encoding; move-to-front encoding; Huffman encoding.
2. Program: GUI built with buttons etc.
3. Program: Efficient Suffix Tree implementation
4. Program: Naive implementation of Burrows-Wheeler
5. Report: How does Burrows-Wheeler work with an emphasis on writing and understanding a proof that the reverse transform works.
6. Report: Shannon information theory and lossy v lossless compression; complexity theory; pseudo-code of algorithms implemented.
7. Reports: Complexity, NP hardness and the big O notation
8. Report: The preprocessing phase for compression.
9. Program: Implement an efficient Burrows-Wheeler Transform using Suffix Trees.
10. Program: Apply the Lyndon factorisation to an input string, and apply the Burrows-Wheeler transform to each of the Lyndon factors.

Final Deliverables

1. The program will have a Graphical User Interface with which to input text files and display compression ratios for compressed files.
2. The Lyndon factorisation and Burrows-Wheeler transformation of a string will be displayed.
3. The program must have a full object-oriented design, with a full implementation life cycle using modern software engineering principles.
4. The report will describe the software engineering process involved in generating your software, interesting programming techniques and data structures used (or able to be used) on the project.
5. The report will describe the complexity analysis of your computations.
6. The report will include some results of benchmarking compression algorithms with a discussion of their meanings.
7. The report will include discussion of applications of the Burrows-Wheeler transform and Lyndon factorisation.
8. The report will include a User Manual.

Suggested Extensions

• Build a mobile Burrow-Wheeler application: this will require use of Worker Threads.
• Show the output using an effective JTree with callbacks to enable user interaction.
• Animate the Burrows-Wheeler in order to help explain its working to a student unfamiliar with the algorithm.
• Adding Digital Signatures to compressed files.

Reading

• D. Adjeroh, T. Bell, A. Mukherjee, The Burrows-Wheeler transform: data compression, suffix arrays, and pattern matching, Springer (2008) 352 pp.
• M. Burrows and D. Wheeler, A block sorting lossless data compression algorithm, Technical Report 124, Digital Equipment Corporation (1994).
• M. Salson, T. Lecroq, M. Leonard and L. Mouchard, A four-stage algorithm for updating a Burrows-Wheeler transform, Theoret. Comput. Sci. 410 (43) (2009) 4350-4359 doi:10.1016/j.tcs.2009.07.016.
• http://en.wikipedia.org/wiki/Burrows-Wheeler transform
• M. Lothaire, Combinatorics onWords, Addison-Wesley, Reading, MA, 1983; 2nd Edition, Cambridge University Press, Cambridge, 1997.
• J. P. Duval, Factorising words over an ordered alphabet, J. Algorithms 4 (1983) 363-381.
• http://en.wikipedia.org/wiki/Data_compression

Early Deliverables

1. A report on the use of context free grammars to specify syntax, along with manual procedures for generating and parsing languages.
2. A report on the use of a parsing tool to produce derivations, and the use of grammar idioms to capture notions of associativity and priority in arithmetic expressions.
3. The development of an interpreter for a language that models a four function calculator with memory.
4. The development of a pretty printer for BNF specifications, demonstrating that programs can process their own source code.
5. A technical report covering all of the above work, including a bibliography.
6. A Term 2 project schedule with a list of deliverables and dates.

Final Deliverables

1. A project report containing a description of all the research you have carried out and the details of your second term implementation.
2. An implementation of your chosen tool, See the suggestions in the Background Section.

Early Deliverables

1. A report on the use of context free grammars to specify syntax, along with manual procedures for generating and parsing languages.
2. A report on the use of a parsing tool to produce derivations, and the use of grammar idioms to capture notions of associativity and priority in arithmetic expressions.
3. The development of an interpreter for a language that models a four function calculator with memory.
4. The development of a pretty printer for BNF specifications, demonstrating that programs can process their own source code.
5. A technical report covering all of the above work, including a bibliography.
6. A Term 2 project schedule with a list of deliverables and dates.

Final Deliverables

1. A project report containing a description of all the research you have carried out and the details of your second term implementation.
2. An implementation of your chosen tool, See the suggestions in the Background Section.

Early Deliverables

1. A report describing the state of the art in automated camera systems.
2. A clear workplan describing the set of tests to be performed, tools to be implemented and classes of techniques to be proposed and studied

Final Deliverables

1. Design and implement a framework for development of new and existing physical security techniques
2. Describe in detail the design of tests for the framework
3. Describe and analyse the features that have been used during the project to detect/stop threats
4. Critically compare the findings with related works

Reading

• OpenCV: https://opencv.org/
• TensorFlow: https://www.tensorflow.org/
• Computer vision: a modern approach

Early Deliverables

1. A review and a report of the approach in the case of classification and regression.
2. A design of the system and its extension for conformal predictors.
3. Implementation of a pilot system and some preliminary experiments.

Final Deliverables

1. Implementation of the system, experiments with various parameters.
2. Experimenting with different underlying ML algorithms.
3. It is desirable that the final system will have a graphical user interface.
4. The final report will describe the theory, the implementation issues, computational experiments with different data sets and parameters.

Suggested Extensions

• Online learning with conformal predictors.

Reading

• V.Vovk, A.Gammerman, G.Shafer. "Algorithmic Learning in a Random World", Springer, 2005.

Early Deliverables

1. Find and choose a data set
2. Implement a relevant version of Conformal Prediction algorithm
3. Impute some mistakes into the data, apply the algorithm, and observe the results.

Final Deliverables

1. Implement a Conformal Prediction with an additional function of error detection.
2. Compare different versions of Conformal Prediction by their efficiency for data correction.
3. Analyse and report the output.

Suggested Extensions

• Propose, justify theoretically, and implement a scheme of multiple error correction.
• Compare how the approach works in different (on-line, batch, cross-validation, inductive) modes.

Reading

• Hedging predictions in machine learning by Alex Gammerman and Vladimir Vovk. Computer Journal 50:151--177 (2007).
• A tutorial on conformal prediction by Glenn Shafer and Vladimir Vovk. Journal of Machine Learning Research 9:371--421 (2008).
• Mushroom Data Set. https://archive.ics.uci.edu/ml/datasets/Mushroom

Early Deliverables

1. Find and choose a data set
2. Implement a relevant version of Conformal Prediction algorithm
3. Impute some mistakes into the data, apply the algorithm, and observe the results.

Final Deliverables

1. Implement a Conformal Prediction with an additional function of error detection.
2. Compare different versions of Conformal Prediction by their efficiency for data correction.
3. Analyse and report the output.

Suggested Extensions

• Propose, justify theoretically, and implement a scheme of multiple error correction.
• Compare how the approach works in different (on-line, batch, cross-validation, inductive) modes.

Reading

• Hedging predictions in machine learning by Alex Gammerman and Vladimir Vovk. Computer Journal 50:151--177 (2007).
• A tutorial on conformal prediction by Glenn Shafer and Vladimir Vovk. Journal of Machine Learning Research 9:371--421 (2008).
• Mushroom Data Set. https://archive.ics.uci.edu/ml/datasets/Mushroom

Early Deliverables

1. Study data sets and identify one relevant for the project, describe it and the structure of its label space.
2. Select a basic machine learning approach applicable to this data.
3. Develop and implement a multi-class version of this approach and get some initial results on the data

Final Deliverables

1. Survey of the data and machine learning techniques.
2. Develop a conformity score functions for data with structured label space.
3. Implement a conformal prediction algorithm applicable to the data.
4. Analysis of the output.

Suggested Extensions

• Compare results with different basic machine learning approaches.
• Compare original label space structure attached to the data against its automatic structuring by means of data analysis.

Reading

• Hedging predictions in machine learning by Alex Gammerman and Vladimir Vovk. Computer Journal 50:151--177 (2007).
• A tutorial on conformal prediction by Glenn Shafer and Vladimir Vovk. Journal of Machine Learning Research 9:371--421 (2008).
• Predicting structured objects with support vector machines by T.Joachims, T.Hofmann, Y.Yue, C.N.Yu. Communications of the ACM 52 (11), 97--104

Early Deliverables

1. Study data sets and identify one relevant for the project, describe it and the structure of its label space.
2. Select a basic machine learning approach applicable to this data.
3. Develop and implement a multi-class version of this approach and get some initial results on the data

Final Deliverables

1. Survey of the data and machine learning techniques.
2. Develop a conformity score functions for data with structured label space.
3. Implement a conformal prediction algorithm applicable to the data.
4. Analysis of the output.

Suggested Extensions

• Compare results with different basic machine learning approaches.
• Compare original label space structure attached to the data against its automatic structuring by means of data analysis.

Reading

• Hedging predictions in machine learning by Alex Gammerman and Vladimir Vovk. Computer Journal 50:151--177 (2007).
• A tutorial on conformal prediction by Glenn Shafer and Vladimir Vovk. Journal of Machine Learning Research 9:371--421 (2008).
• Predicting structured objects with support vector machines by T.Joachims, T.Hofmann, Y.Yue, C.N.Yu. Communications of the ACM 52 (11), 97--104

Early Deliverables

1. You will implement simple inductive conformal predictors and run them on small artificial data sets.
2. They will be based both on underlying algorithms implemented by you (such as nearest neighbours)
3. ... and on advanced machine learning algorithms included in available libraries (such as gradient boosting).
4. You will produce a report containing descriptions of algorithms and basics of the theory.

Final Deliverables

1. You will implement transductive conformal predictors based on the method of k nearest neighbours; transductive conformal predictors are less computationally efficient but have better accuracy.
2. You will implement cross-conformal predictors, which combine the accuracy of transductive conformal predictors and the computational efficiency of inductive conformal predictors.
3. The algorithms will be studied empirically on benchmark data sets such as those available from the UCI data repository and the Delve repository.
4. For many of these data set judicious preprocessing (including normalisation of attributes or examples) will be essential.
5. Two aspects of the performance of the algorithms will be explored: their validity and efficiency.
6. The overall program will have a full object-oriented design, with a full implementation life cycle using modern software engineering principles.
7. Ideally, it will have a graphical user interface.
8. The report will describe: the theory behind the algorithms.
9. The report will describe: the implementation issues necessary to apply the theory.
10. The report will describe: the software engineering process involved in generating your software.
11. The report will describe: computational experiments with different data sets, underlying algorithms, and parameters.

Suggested Extensions

• Empirical study of the performance of the method of cross validation for different numbers of folds and extend to the method of cross-conformal prediction.
• Exploring the validity of conformal predictors conditional on the training set.
• Studying connections between probability-type inductive conformal predictors and ROC curves.

Reading

• Glenn Shafer and Vladimir Vovk. A tutorial on conformal prediction. Journal of Machine Learning Research, 9:371-421, 2008. http://arxiv.org/abs/0706.3188
• Alex Gammerman and Vladimir Vovk. Hedging predictions in machine learning (with discussion). The Computer Journal, 50:151-177, 2007. http://arxiv.org/abs/cs/0611011
• Vladimir Vovk, Alex Gammerman, and Glenn Shafer. Algorithmic Learning in a Random World. Springer, New York, 2005.
• Vladimir Vovk. Cross-conformal predictors. http://arxiv.org/abs/1208.0806
• Vladimir Vovk. Inductive conformal predictors in the batch mode. http://arxiv.org/abs/1209.2673
• Seth Hettich and Steven D. Bay. The UCI KDD Archive. University of California, Department of Information and Computer Science, Irvine, CA, 1999, http://kdd.ics.uci.edu
• Delve archive. http://www.cs.toronto.edu/~delve

Early Deliverables

1. Producing a simple proof-of-concept implementation of classification-based conformal prediction with different underlying algorithms (e.g. SVM, k-NN, Random Forests, etc.) from scratch.
2. A report of the empirical results of such conformal predictors on a small synthetic dataset.

Final Deliverables

1. A detailed literature review of existing machine learning algorithms used for non-speech sound recognition.
2. Producing a detailed report of the dataset structure, including some visual analysis (e.g. t-SNE, UMAP, etc.).
3. Implementing conformal predictors on the whole dataset, optimising the underlying algorithms' parameters.
4. Carefully analysing the p-values, comparing the validity, efficiency, accuracy of the implemented conformal predictors.

Suggested Extensions

• Comparing different methods of extracting sound features (e.g. MFCC, CFA, ZCR) from raw recordings, and their impacts on the prediction performance.
• Using the entire dataset to train a new model, then testing such model on other real world recordings.

Reading

• Nguyen, K. A. and Luo, Z. (2018). Cover Your Cough: Detection of Respiratory Events with Confidence Using a Smartwatch. Proceedings of Machine Learning Research, 91, 1-18.
• Piczak, K. J. (2015). ESC: Dataset for environmental sound classification. In Proceedings of the 23rd ACM international conference on Multimedia (pp. 1015-1018). ACM.
• https://dataverse.harvard.edu/dataset.xhtml?persistentId=doi:10.7910/DVN/YDEPUT

Early Deliverables

1. Producing a simple proof-of-concept implementation of classification-based conformal prediction with different underlying algorithms (e.g. SVM, k-NN, Random Forests, etc.) from scratch.
2. A report of the empirical results of such conformal predictors on a small synthetic dataset.

Final Deliverables

1. A detailed literature review of existing machine learning algorithms used for non-speech sound recognition.
2. Producing a detailed report of the dataset structure, including some visual analysis (e.g. t-SNE, UMAP, etc.).
3. Implementing conformal predictors on the whole dataset, optimising the underlying algorithms' parameters.
4. Carefully analysing the p-values, comparing the validity, efficiency, accuracy of the implemented conformal predictors.

Suggested Extensions

• Comparing different methods of extracting sound features (e.g. MFCC, CFA, ZCR) from raw recordings, and their impacts on the prediction performance.
• Using the entire dataset to train a new model, then testing such model on other real world recordings.

Reading

• Nguyen, K. A. and Luo, Z. (2018). Cover Your Cough: Detection of Respiratory Events with Confidence Using a Smartwatch. Proceedings of Machine Learning Research, 91, 1-18.
• Piczak, K. J. (2015). ESC: Dataset for environmental sound classification. In Proceedings of the 23rd ACM international conference on Multimedia (pp. 1015-1018). ACM.
• https://dataverse.harvard.edu/dataset.xhtml?persistentId=doi:10.7910/DVN/YDEPUT

Early Deliverables

1. A report on Hidden Markov Model (HMM) tools in Bioinformatics.
2. A report on conformal predictors.
3. A simple implementation of conformal prediction on a family of proteins from PFAM.

Final Deliverables

1. A framework to evaluate the credibility and confidence of individual HMM's from PFAM.

Suggested Extensions

• Using conformal predictors on HMM models in protein sequence prediction.

Reading

• The Pfam protein families database. Finn RD1, Mistry J, Tate J, Coggill P, Heger A, Pollington JE, Gavin OL, Gunasekaran P, Ceric G, Forslund K, Holm L, Sonnhammer EL, Eddy SR, Bateman A. doi: 10.1093/nar/gkp985.
• http://jmlr.csail.mit.edu/papers/volume9/shafer08a/shafer08a.pdf

Early Deliverables

1. A report on Hidden Markov Model (HMM) tools in Bioinformatics.
2. A report on conformal predictors.
3. A simple implementation of conformal prediction on a family of proteins from PFAM.

Final Deliverables

1. A framework to evaluate the credibility and confidence of individual HMM's from PFAM.

Suggested Extensions

• Using conformal predictors on HMM models in protein sequence prediction.

Reading

• The Pfam protein families database. Finn RD1, Mistry J, Tate J, Coggill P, Heger A, Pollington JE, Gavin OL, Gunasekaran P, Ceric G, Forslund K, Holm L, Sonnhammer EL, Eddy SR, Bateman A. doi: 10.1093/nar/gkp985.
• http://jmlr.csail.mit.edu/papers/volume9/shafer08a/shafer08a.pdf

Early Deliverables

1. Familiarisation with the idea of strategies and the identification of set of them.
2. Implementation of strategies and the definition of an interface that allows agents to be selected and play against each other
3. Definition of a tournament
4. Report on the design of a tournament.
5. Implementation of a tournament, where a user selects the number of agents, associates them with existing strategies, specifies the number of iterations and deploys a tournament
6. Report on the Iterative Prisoner's dilemma problem
7. Report on previous work with the Iterative Prisoner's dilemma

Final Deliverables

1. A program that allows us to simulate tournaments of the iterative prisoner's dilemma.
2. A GUI that allows a user to create and manage tournaments.
3. GUI should provide summary statistics that are tournament specific, i.e. a summary for each player/strategy and how it is compared with the other players/strategies in the tournament.
4. The report should describe some of the theory behind strategies in cooperative settings.
5. The report should provide an analysis and an evaluation of the different strategies, including an indication of which one is the best strategy, if any.
6. The report should contain a design of the program and a discussion of any software engineering issues encountered.
7. The report should describe any useful and interesting programming techniques that have been employed to develop the final prototype.

Suggested Extensions

• Make use of multi-agent system platform to support the tournament in a completely distributed setting.
• Use logic-based rules to specify the strategies and wrap them in the Java implementation.
• Use strategy recognition techniques so that agents can detect what strategies opponents are playing, and adapt their behaviour accordingly.
• Create a methodology for developing tournaments of this type.

Reading

• M. Nowak, R. May and K. Sigmund. The Arithmetics of Mutual Help, Scientific American, June 1995.
• R. Axelrod, W. D. Hamilton (1981), The Evolution of Cooperation, Science 211: 1390–96.
• K. Stathis, A game-based architecture for developing interactive components in computational logic, Journal of Functional and Logic Programming, 20(1), Jan 2000, MIT Press.

Early Deliverables

1. A proof-of-concept program implementing neural nets both in the case of regression and in the case of classification (where neural networks output probability predictions).
2. Building a cross-conformal predictor on top of neural nets and exploring its empirical validity.
3. A preliminary empirical study of the efficiency of the student's implementation and standard implementations (such as the one in scikit-learn).

Final Deliverables

1. Careful cross-validation of various parameters, including the stopping rule, learning rate, and the numbers of hidden neurons at various levels.
2. Implementing regularization.
3. Experimenting with different architectures, possibly including deep architectures.
4. Ideally, the final product will have a graphical user interface.
5. The final report will describe the theory behind the algorithms, the implementation issues necessary to apply the theory, the software engineering process involved in generating your software, computational experiments with different data sets, methods, and parameters.

Suggested Extensions

• Careful study of the optimal number of folds.
• Replacing back-propagation with other optimization algorithms.
• Implementing jackknife+, a popular recent version of cross-conformal prediction.

Reading

• Hastie, Tibshirani, and Friedman, The Elements of Statistical Learning, 2009, Chapter 11.
• Mitchell, Machine Learning, 1997, Chapter 4.
• Research papers.

Early Deliverables

1. A proof-of-concept program implementing neural nets both in the case of regression and in the case of classification (where neural networks output probability predictions).
2. Building a cross-conformal predictor on top of neural nets and exploring its empirical validity.
3. A preliminary empirical study of the efficiency of the student's implementation and standard implementations (such as the one in scikit-learn).

Final Deliverables

1. Careful cross-validation of various parameters, including the stopping rule, learning rate, and the numbers of hidden neurons at various levels.
2. Implementing regularization.
3. Experimenting with different architectures, possibly including deep architectures.
4. Ideally, the final product will have a graphical user interface.
5. The final report will describe the theory behind the algorithms, the implementation issues necessary to apply the theory, the software engineering process involved in generating your software, computational experiments with different data sets, methods, and parameters.

Suggested Extensions

• Careful study of the optimal number of folds.
• Replacing back-propagation with other optimization algorithms.
• Implementing jackknife+, a popular recent version of cross-conformal prediction.

Reading

• Hastie, Tibshirani, and Friedman, The Elements of Statistical Learning, 2009, Chapter 11.
• Mitchell, Machine Learning, 1997, Chapter 4.
• Research papers.

Early Deliverables

1. A report describing the state of the art in resilience.
2. A clear workplan describing the set of tests to be performed, tools to be implemented and classes of techniques to be proposed and studied.

Final Deliverables

1. Design and implement a framework for development of new and existing visualization techniques.
2. Describe in detail the design of tests for the framework.
3. Describe and analyse the features that have been used during the project to detect/stop threats.
4. Critically compare the findings with related works.

Reading

• WebGL fundamentals https://webgl2fundamentals.org/
• SecViz.org https://secviz.org/
• Bokeh, Vega, D3, Plotly visualization suites
• R Marty. Applied security visualization.
• H Shiravi et al. A survey of visualization systems for network security.

Early Deliverables

1. Proof of concept program: Implementation of a deep learning algorithm for a relatively small network on a relatively simple dataset, together with experiments determining the effect of hyperparameters on convergence
2. A report giving an overview of a range of related deep learning methods.
3. A report describing and assessing the performance of the initial implementation and any experiments performed

Final Deliverables

1. The program(s) will work with a real dataset(s).
2. You will implement and compare multiple architectures and optimisation methods
3. You will visualize and assess the performance of the algorithms produced.
4. The report will describe and derive the formulae for the methods that you use, and the report will also discuss the implementation.

Suggested Extensions

• This would be an opportunity to implement several deep learning methods and get to understand them in practice by applying them to a significant dataset.

Reading

• "Deep Learning", by Goodfellow, Bengio, and Courville, MIT Press 2016
• "Deep Learning with Python" by Francois Chollet, Manning Publications 2018

Early Deliverables

1. Proof of concept program: Implementation of a deep learning algorithm for a relatively small network on a relatively simple dataset, together with experiments determining the effect of hyperparameters on convergence
2. A report giving an overview of a range of related deep learning methods.
3. A report describing and assessing the performance of the initial implementation and any experiments performed

Final Deliverables

1. The program(s) will work with a real dataset(s).
2. You will implement and compare multiple architectures and optimisation methods
3. You will visualize and assess the performance of the algorithms produced.
4. The report will describe and derive the formulae for the methods that you use, and the report will also discuss the implementation.

Suggested Extensions

• This would be an opportunity to implement several deep learning methods and get to understand them in practice by applying them to a significant dataset.

Reading

• "Deep Learning", by Goodfellow, Bengio, and Courville, MIT Press 2016
• "Deep Learning with Python" by Francois Chollet, Manning Publications 2018

Early Deliverables

1. A report describing the state of the art in resilience.
2. A clear workplan describing the set of tests to be performed, tools to be implemented and classes of techniques to be proposed and studied

Final Deliverables

1. Design and implement a framework for development of new and existing resilience techniques
2. Describe in detail the design of tests for the framework
3. Describe and analyse the features that have been used during the project to detect/stop threats
4. Critically compare the findings with related works

Reading

• Julia Allen. Measures for managing operational resilience.Technical Report, 2011.
• Julia Allen, Pamela Curtis, Nader Mehravari, Andrew Moore, Kevin Partridge, Robert Stoddard, and Randy Trzeciak. Analyzing cases of resilience success and failure-a research study. Technical report, Carnegie Mellon University, the Software Engineering Institute, 2012.
• Richard A Caralli, Julia Allen, and David W White.CERT resilience management model: A maturity model for managing operational resilience. Addison-Wesley Professional, 2010.
• Deborah Bodeau and Richard Graubart. Cyber resilience metrics: Key observations. Technical Report,2016.
• Ronald J Brachman, Richard E Fikes, and Hector J Levesque. Krypton: A functional approach to knowledge representation.Computer, (10):67–73, 1983.
• Linkov and Trump. The science and practice of resilience. 2019

Early Deliverables

1. A report describing the state of the art in malware disguising themselves as other processes.
2. A clear workplan describing the set of tests to be performed, tools to be implemented and classes of techniques to be proposed and studied

Final Deliverables

1. Design and implement a framework for development of new and existing related malware detection techniques
2. Describe in detail the design of tests for the framework
3. Describe and analyse the features that have been used during the project to detect/stop threats
4. Critically compare the findings with related works

Reading

• Vaas and Happa. Detecting disguised processes using application-behavior profiling. 2017
• Geden and Happa. Classification of malware families based on runtime behaviour. 2018

Early Deliverables

1. A report describing different security apps for integrating into security suite.
2. A proof-of-concept implementation of the security suite
3. Testing the system

Final Deliverables

1. Design and develop the final security suite
2. Complete report which must include a User Manual

Reading

• Conti, M et al. Context-related policy enforcement for android
• Bojjagani, S. et al. Stamba: Security testing for Android mobile banking apps.
• Vulnerability Test Suite for Android - NowSecure https://info.nowsecure.com › VTS-for-Android
• https://www.android.com/intl/en_uk/enterprise/security/

Early Deliverables

1. Implement a user-friendly signing tool that applies a digital signature to a given file
2. Implement a TSA service that can be accessed via a REST API
3. Design a suitable file format to accumulate digital evidence

Final Deliverables

1. Extend the signing tool with capabilities to automatically interact with TSAs to generate and refresh digital evidence
2. Extend the signing tool with capabilities to verify whether a digitally signed and timestamped document existed at a certain point in time

Reading

• Martín Vigil, Johannes Buchmann, Daniel Cabarcas, Christian Weinert, Alexander Wiesmaier: "Integrity, Authenticity, Non-Repudiation, and Proof of Existence for Long-Term Archiving: A Survey". In Computers & Security 2015.
• Christian Weinert, Denise Demirel, Martín Vigil, Matthias Geihs, Johannes Buchmann: "MoPS: A Modular Protection Scheme for Long-Term Storage". In ASIACCS 2017.

Early Deliverables

1. Algorithms for three new insertion strategies with different trade-offs (in terms of time and storage complexity)
2. Implementation of at least one of the new strategies in one existing Cuckoo filter implementation
3. Implementation of a simulation environment that will allow to measure the success and run-time overhead of different insertion strategies

Final Deliverables

1. Implementation of the remaining new insertion strategies
2. Simulation results for all insertion strategies

Reading

• Bin Fan, David G. Andersen, Michael Kaminsky, Michael Mitzenmacher: "Cuckoo Filter: Practically Better Than Bloom". In: CoNEXT 2014.
• Daniel Kales, Christian Rechberger, Thomas Schneider, Matthias Senker, Christian Weinert: "Mobile Private Contact Discovery at Scale". In USENIX Security Symposium 2019.
• https://github.com/efficient/cuckoofilter
• https://github.com/linvon/cuckoo-filter
• https://github.com/MGunlogson/CuckooFilter4J
• https://github.com/axiomhq/rust-cuckoofilter

Early Deliverables

1. A basic implementation for sending sequences of characters. The 'droplets' may be object instances with appropriate fields. The implementation should consist of two programs: an encoder that reads in a text file and writes out a sequence of 'droplets', and a decoder that reads in a sequence of droplets and which outputs a text file, which should be identical to the original text file.
2. A description of how the basic encoding and decoding algorithms work, with pseudocode and a running example. It should also include a discussion of the 'degree distribution', and of how to sample random values from a discrete distribution.

Final Deliverables

1. An efficient program for encoding and decoding.
2. A sequence of well designed and analysed experiments investigating the program's efficiency and behaviour.
3. The report will contain a discussion of erasure channels, contrast with feedback channels. When are erasure channels needed? What other solutions are there?
4. The report will contain a description of Digital Fountain Codes: basic encoding and decoding, with a running example. Role of degree distribution. Behaviour with special cases of degree distribution. Luby's degree distributions. Time and space efficiency of variants of decoding algorithm.
5. The report will contain 3. Design and implementation: What is overall structure of program? What are design issues? How were these design issues resolved?
6. The report will contain complete testing schedule: What is evidence program is correct (test cases)
7. The report will contain an analysis of the time and space efficiency of algorithm(s)
8. The report will contain, for each experiment: aim, method, tables/graphs of results, questions raised

Suggested Extensions

• Implementation should be efficient: different optimisations lead to time and space efficiency. Many small extensions to algorithm are possible. You could design and carry out experiments to assess the efficiency of your algorithm.
• It is possible to design experiments to check the efficiency of the packet degree distribution, and you could try to optimise the degree distribution in the light of your experiments.

Reading

• Start with Wikipedia on 'Luby Transform Code', and the references it gives.

Early Deliverables

1. A basic implementation for sending sequences of characters. The 'droplets' may be object instances with appropriate fields. The implementation should consist of two programs: an encoder that reads in a text file and writes out a sequence of 'droplets', and a decoder that reads in a sequence of droplets and which outputs a text file, which should be identical to the original text file.
2. A description of how the basic encoding and decoding algorithms work, with pseudocode and a running example. It should also include a discussion of the 'degree distribution', and of how to sample random values from a discrete distribution.

Final Deliverables

1. An efficient program for encoding and decoding.
2. A sequence of well designed and analysed experiments investigating the program's efficiency and behaviour.
3. The report will contain a discussion of erasure channels, contrast with feedback channels. When are erasure channels needed? What other solutions are there?
4. The report will contain a description of Digital Fountain Codes: basic encoding and decoding, with a running example. Role of degree distribution. Behaviour with special cases of degree distribution. Luby's degree distributions. Time and space efficiency of variants of decoding algorithm.
5. The report will contain 3. Design and implementation: What is overall structure of program? What are design issues? How were these design issues resolved?
6. The report will contain complete testing schedule: What is evidence program is correct (test cases)
7. The report will contain an analysis of the time and space efficiency of algorithm(s)
8. The report will contain, for each experiment: aim, method, tables/graphs of results, questions raised

Suggested Extensions

• Implementation should be efficient: different optimisations lead to time and space efficiency. Many small extensions to algorithm are possible. You could design and carry out experiments to assess the efficiency of your algorithm.
• It is possible to design experiments to check the efficiency of the packet degree distribution, and you could try to optimise the degree distribution in the light of your experiments.

Reading

• Start with Wikipedia on 'Luby Transform Code', and the references it gives.

Early Deliverables

1. Familiarity with necessary background on Intel SGX.
2. Familiarity with Intel SGX SDK. Show examples of its operation on toy/tutorial problems.
3. Report: Intel SGX overview
4. Report: Distributed Games, including architecture and potential for cheating in the game of your choice.
5. Proof-of-concept: Partition a real-world game of your choice (e.g. ioquake3) so that the server runs inside an enclave
6. Proof-of-concept: Offload some of the computation at the server to the client, inside an enclave

Final Deliverables

1. Modify the communication protocol between the client and the server so that the integrity of the game is preserved even when a client is cheating;
2. Establish a secure communication protocol, e.g., TLS, so that an attacker cannot observe the network traffic for cheating;
3. Evaluate the performance and scalability benefits of your distributed game.
4. Report: Detailed architecture description of final partitioned game.
5. Report: Security analysis of final partitioned game.

Suggested Extensions

• Using the Intel SGX functionalities, establish a framework to remotely attest/verify the authenticity of a client/server and provision it with sensitive data.

Reading

• https://support.steampowered.com/kb_article.php?ref=7849-Radz-6869
• http://www.gamesradar.com/diablo-iii-error-37-causing-great-wailing-and-gnashing-teeth-players-worldwide-fail-log-solution-wait-a-bit/
• http://www.ibtimes.com/division-servers-struggle-ubisoft-confirms-login-problems-2332172
• Raaen et al., Latency Thresholds for Usability in Games: A Survey, NIK 2014.
• Howard et al., Cascading Impact of Lag on User Experience in Cooperative Multiplayer Games, NetGames 2014.
• https://software.intel.com/en-us/sgx
• https://ioquake3.org/
• Yahyavi et al., Watchmen: Scalable cheat-resistant support for distributed multi-player online games, ICDCS 2013.
• Intel SGX Linux SDK Sample Code: https://github.com/intel/linux-sgx/tree/master/SampleCode/SampleEnclave
• Intel SGX Explained, https://eprint.iacr.org/2016/086.pdf

Early Deliverables

1. A report describing architectural components of the IoT system and their relevant activities which need to be audited and addressing Blockchain technologies available for use
2. A design of the auditing data model
3. A proof of concept implementation of the auditing system using a DLT technologies
4. Testing the system

Final Deliverables

1. Design and develop the data model of the IoT-based auditing system
2. Develop a prototype of Blockchain-IoT-based auditing system
3. Complete report
4. The report will include a User Manual

Reading

• https://www.networkmanagementsoftware.com/what-is-syslog/
• https://en.wikipedia.org/wiki/Syslog
• https://medium.com/swlh/top-5-dlt-blockchain-protocols-to-consider-in-your-project-dae7fc6d381d
• https://www.ibm.com/blogs/research/2018/02/architecture-hyperledger-fabric/
• https://hyperledger-fabric.readthedocs.io/en/release-1.4/getting_started.html

Early Deliverables

1. A proof-of-concept implementation of bagging, random forests, or boosting both in the case of regression and in the case of classification.
2. Implementing underlying algorithms, such as decision stumps or decision trees; simple proof-of-concept implementations are sufficient.

Final Deliverables

1. A final program implementing bagging, random forests, or boosting (or all three) with a careful choice of the parameters.
2. Implementations of the underlying algorithms should be complemented by a careful empirical study of the effect of various parameters; producing optimized programs.
3. Comparing the performance of different ensemble techniques in the problems of classification and regression.
4. Using bagged and boosted decision trees as probability predictors; evaluating their perfomance using proper loss functions.
5. Ideally, the final product will have a graphical user interface.
6. The final report will describe the theory behind the algorithms, the implementation issues necessary to apply the theory, the software engineering process involved in generating your software, computational experiments with different data sets, methods, and parameters.

Suggested Extensions

• Implementing the out-of-bag estimate of error probability.
• Investigating variable importance.
• Proximity plots.
• Theoretical analysis of the algorithms.

Reading

• Hastie, Tibshirani, and Friedman, The Elements of Statistical Learning, 2009, Section 8.7, Chapters 10 and 15-16
• James, Witten, Hastie, and Tibshirani, An Introduction to Statistical Learning, 2013, Chapter 8.
• Research papers.

Early Deliverables

1. A proof-of-concept implementation of bagging, random forests, or boosting both in the case of regression and in the case of classification.
2. Implementing underlying algorithms, such as decision stumps or decision trees; simple proof-of-concept implementations are sufficient.

Final Deliverables

1. A final program implementing bagging, random forests, or boosting (or all three) with a careful choice of the parameters.
2. Implementations of the underlying algorithms should be complemented by a careful empirical study of the effect of various parameters; producing optimized programs.
3. Comparing the performance of different ensemble techniques in the problems of classification and regression.
4. Using bagged and boosted decision trees as probability predictors; evaluating their perfomance using proper loss functions.
5. Ideally, the final product will have a graphical user interface.
6. The final report will describe the theory behind the algorithms, the implementation issues necessary to apply the theory, the software engineering process involved in generating your software, computational experiments with different data sets, methods, and parameters.

Suggested Extensions

• Implementing the out-of-bag estimate of error probability.
• Investigating variable importance.
• Proximity plots.
• Theoretical analysis of the algorithms.

Reading

• Hastie, Tibshirani, and Friedman, The Elements of Statistical Learning, 2009, Section 8.7, Chapters 10 and 15-16
• James, Witten, Hastie, and Tibshirani, An Introduction to Statistical Learning, 2013, Chapter 8.
• Research papers.

Early Deliverables

1. A report on graphs and scale-free networks.
2. A report on the PID graph.
3. A report on GraphQL.

Final Deliverables

1. The development of a tool to query the PID graph and visualise graphs.
2. Identifying the behaviour of the degree of the PID graph.

Suggested Extensions

• Exploring other graph obseravbles e.g. centrality.
• Exploring using a recommender system to identify papers of interest.

Reading

• Introducing the PID Graph - https://blog.datacite.org/introducing-the-pid-graph/
• The Fullstack Introduction for GraphQL - https://www.howtographql.com/
• Learning GraphQL, E.Porcello and A. Banks O'Reilly Media (available online through Safari Books)
• Handbook of Graph Theory, J.L. Gross, J. Yellen and P. Zhang, Chapman and Hall/C (available online through Safari Books)
• Scale-free networks are rare, A.D. Broido and A. Clauset, Nature Comms. (2019) DOI - 10.1038/s41467-019-08746-5

Early Deliverables

1. A report on graphs and scale-free networks.
2. A report on the PID graph.
3. A report on GraphQL.

Final Deliverables

1. The development of a tool to query the PID graph and visualise graphs.
2. Identifying the behaviour of the degree of the PID graph.

Suggested Extensions

• Exploring other graph obseravbles e.g. centrality.
• Exploring using a recommender system to identify papers of interest.

Reading

• Introducing the PID Graph - https://blog.datacite.org/introducing-the-pid-graph/
• The Fullstack Introduction for GraphQL - https://www.howtographql.com/
• Learning GraphQL, E.Porcello and A. Banks O'Reilly Media (available online through Safari Books)
• Handbook of Graph Theory, J.L. Gross, J. Yellen and P. Zhang, Chapman and Hall/C (available online through Safari Books)
• Scale-free networks are rare, A.D. Broido and A. Clauset, Nature Comms. (2019) DOI - 10.1038/s41467-019-08746-5

Early Deliverables

1. Proof of concept programs: bandit problem solver ('bandit problems' are particularly trivial statistical games, which you should tackle first), with experiments showing how upper-confidence-bound (UCB) methods compare to naive strategies.
2. Proof of concept programs: Program that enables 2 human players to play the selected game (no AI moves).
3. Reports: UCB method for bandit problems.
4. Reports: Rules of the game, expressed semi-formally as state and allowable transitions.
5. Reports: Algorithm for MCTS.

Final Deliverables

1. The program must have a full object-oriented design, with a full implementation life cycle using modern software engineering principles
2. The program should have a GUI for game play, for the case of two human players, and one player against the computer.
3. The report should include experimental assessments of effects on program performance of parameters of MCTS (depth of search, number of rollouts used, etc).
4. The report will describe the software engineering process involved in generating your software.
5. The report will have sections on: the theory of bandit processes, and the algorithm for MCTS; the game used, the software engineering process in development of the program; an explanation and justification of design decisions for the GUI; and an experimental analysis of the performance of MCTS.

Suggested Extensions

• A possible enhancement would be a comparison of MCTS with alpha-beta pruning.
• The project could be done for a relatively simple game ( such as Connect4 ) or for 9x9 capture Go, which adds some programming complexity.

Early Deliverables

1. Proof of concept programs: First implementation of GA for constraint satisfaction/function optimisation.
2. Report on Design Patterns.
3. Report on Genetic Algorithms.
4. Report on encoding various problems for GAs.
5. Report on the Theory of coalescence and genetic drift.

Final Deliverables

1. The program must have a full object-oriented design, with a full implementation life cycle using modern software engineering principles
2. The program will at least enable a user to interactively view and control GA optimisation for (at least one) optimisation problem.
3. The program must have a text based control file from which it can be configured.
4. The program will have a Graphical User Interface that can be used to animate simulations.
5. The report will describe the software engineering process involved in generating your software.
6. The report will include a survey of optimisation methods, including GAs.
7. The report will include some results of experiments that compare different optimisation methods applied to the same problems.
8. The report will describe interesting programming techniques and data structures used (or able to be used) on the project.

Suggested Extensions

• This is an open-ended project that requires a substantial amount of programming. The basic project may be extended in many ways.
• One possibility would be to choose an optimisation problem, and then compare a range of optimisation techniques for it, including GAs.

Early Deliverables

1. Proof of concept program: identify one machine learning task suitable for characters recognition and implement one learning algorithm for the chosen task.
2. Report: An overview of the datasets. Statistics and numbers involved.
3. Report: A report on machine learning tasks for character recognition.

Final Deliverables

1. Survey of machine learning techniques used for characters recognition.
2. The developed machine learning software for Greek characters recognition.
3. The software engineering process involved in generating your software.

Suggested Extensions

• Provide predictions with confidence measure.

Reading

• Gatos, Basilis, et al. GRPOLY-DB: An old Greek polytonic document image database. Document Analysis and Recognition (ICDAR), 2015 13th International Conference on. IEEE, 2015.
• http://www.iit.demokritos.gr/~nstam/GRPOLY-DB

Early Deliverables

1. Proof of concept program: identify one machine learning task suitable for characters recognition and implement one learning algorithm for the chosen task.
2. Report: An overview of the datasets. Statistics and numbers involved.
3. Report: A report on machine learning tasks for character recognition.

Final Deliverables

1. Survey of machine learning techniques used for characters recognition.
2. The developed machine learning software for Greek characters recognition.
3. The software engineering process involved in generating your software.

Suggested Extensions

• Provide predictions with confidence measure.

Reading

• Gatos, Basilis, et al. GRPOLY-DB: An old Greek polytonic document image database. Document Analysis and Recognition (ICDAR), 2015 13th International Conference on. IEEE, 2015.
• http://www.iit.demokritos.gr/~nstam/GRPOLY-DB

Early Deliverables

1. Proof of concept program pre-processing the data and applying two machine learning methods to Boston Housing dataset
2. Report: A report describing supervised learning, the concepts 'training set' and 'test set', and the theory of the methods.

Final Deliverables

1. A machine learning method (such as Ridge Regression; the choice is by discussion with the supervisor) should be implemented from scratch.
2. Several different datasets should be used
3. Tests comparing different kernels, parameters, etc should be performed. The results should be visualised.
4. Prediction in the online mode should be implemented and results visuaised.
5. The report will describe different algorithms and implementation issues (such as the choice of data structures, numerical methods etc) necessary to apply the theory.
6. The report will describe computational experiments, analyse their results and draw conclusions

Suggested Extensions

• Methods of prediction with expert advice can be applied in the online mode.
• Is there inflation? A study can be undertaken.

Reading

• Boston housing dataset on Kaggle: https://www.kaggle.com/prasadperera/the-boston-housing-dataset
• California housing dataset on Kaggle: https://www.kaggle.com/camnugent/california-housing-prices
• De Cock, Dean. Ames, Iowa: Alternative to the Boston housing data as an end of semester regression project. Journal of Statistics Education 19.3 (2011).
• Y.Kalnishkan, Kernel Methods, http://onlineprediction.net/?n=Main.KernelMethods
• J.Shawe-Taylor and N.Cristianini, Kernel Methods for Pattern Analysis Cambridge University Press, 2004
• B.Schlkopf and A.J.Smola, Learning with Kernels: Support Vector Machines, Regularization, Optimization, and Beyond, The MIT Press, 2001.

Early Deliverables

1. Proof of concept program pre-processing the data and applying two machine learning methods to Boston Housing dataset
2. Report: A report describing supervised learning, the concepts 'training set' and 'test set', and the theory of the methods.

Final Deliverables

1. A machine learning method (such as Ridge Regression; the choice is by discussion with the supervisor) should be implemented from scratch.
2. Several different datasets should be used
3. Tests comparing different kernels, parameters, etc should be performed. The results should be visualised.
4. Prediction in the online mode should be implemented and results visuaised.
5. The report will describe different algorithms and implementation issues (such as the choice of data structures, numerical methods etc) necessary to apply the theory.
6. The report will describe computational experiments, analyse their results and draw conclusions

Suggested Extensions

• Methods of prediction with expert advice can be applied in the online mode.
• Is there inflation? A study can be undertaken.

Reading

• Boston housing dataset on Kaggle: https://www.kaggle.com/prasadperera/the-boston-housing-dataset
• California housing dataset on Kaggle: https://www.kaggle.com/camnugent/california-housing-prices
• De Cock, Dean. Ames, Iowa: Alternative to the Boston housing data as an end of semester regression project. Journal of Statistics Education 19.3 (2011).
• Y.Kalnishkan, Kernel Methods, http://onlineprediction.net/?n=Main.KernelMethods
• J.Shawe-Taylor and N.Cristianini, Kernel Methods for Pattern Analysis Cambridge University Press, 2004
• B.Schlkopf and A.J.Smola, Learning with Kernels: Support Vector Machines, Regularization, Optimization, and Beyond, The MIT Press, 2001.

Early Deliverables

1. A report on Gene Ontologies.
2. A report on the tool GOCat explaining the methods.
3. A report on the data set required to build the tool.

Final Deliverables

1. The development of a stand alone tool to implement the equivalent of GOCat.
2. An API to query the tool.
3. Rigourous testing of the tool.

Suggested Extensions

• Using more advanced methods beyond kNN for the implementation.
• An API to query the tool.

Reading

• Managing the data deluge: data-driven GO category assignment improves while complexity of functional annotation increases. Gobeill J Pasche E Vishnyakova D Ruch P DOI 10.1093/database/bat041
• A survey on annotation tools for the biomedical literature. Neves M Leser U DOI 10.1093/bib/bbs084
• The GOA database in 2009--an integrated Gene Ontology Annotation resource Barrell D Dimmer E Huntley R Binns D O'Donovan C Apweiler R dx.doi.org/10.1093/nar/gkn803

Early Deliverables

1. A report on Gene Ontologies.
2. A report on the tool GOCat explaining the methods.
3. A report on the data set required to build the tool.

Final Deliverables

1. The development of a stand alone tool to implement the equivalent of GOCat.
2. An API to query the tool.
3. Rigourous testing of the tool.

Suggested Extensions

• Using more advanced methods beyond kNN for the implementation.
• An API to query the tool.

Reading

• Managing the data deluge: data-driven GO category assignment improves while complexity of functional annotation increases. Gobeill J Pasche E Vishnyakova D Ruch P DOI 10.1093/database/bat041
• A survey on annotation tools for the biomedical literature. Neves M Leser U DOI 10.1093/bib/bbs084
• The GOA database in 2009--an integrated Gene Ontology Annotation resource Barrell D Dimmer E Huntley R Binns D O'Donovan C Apweiler R dx.doi.org/10.1093/nar/gkn803

Early Deliverables

1. Implementation of transfer learning algorithms based on pre-trained networks for any of the following tasks, (1) object detection/classification, (2) image segmentation, and (3) video classification.
2. A report describing transfer learning algorithms implemented with some evaluation results.

Final Deliverables

1. Implementing transfer learning, deep learning (e.g. Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs)), or hybrid deep networks for any of the aforementioned tasks.
2. Implementing several deep learning algorithms with different architectures, where the employed architectures are either existing, or newly proposed.
3. Implementing manual hyper-parameter fine-tuning of the implemented deep networks.
4. A comprehensive evaluation and comparison against existing state-of-the-art studies.
5. The final report will describe methodology, implementation and testing with theoretical analysis and detailed evaluation results.

Suggested Extensions

• Implementing deep hybrid networks with new architectures.
• Implementing optimal hyper-parameter selection using an automatic process.
• Theoretical analysis of the algorithms.

Reading

• Goodfellow, I., Bengio, Y. and Courville, A., 2016. Deep learning. MIT press.
• Chollet, F., 2021. Deep learning with Python.
• Journal/conference papers.

Early Deliverables

1. Implementation of transfer learning algorithms based on pre-trained networks for any of the following tasks, (1) object detection/classification, (2) image segmentation, and (3) video classification.
2. A report describing transfer learning algorithms implemented with some evaluation results.

Final Deliverables

1. Implementing transfer learning, deep learning (e.g. Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs)), or hybrid deep networks for any of the aforementioned tasks.
2. Implementing several deep learning algorithms with different architectures, where the employed architectures are either existing, or newly proposed.
3. Implementing manual hyper-parameter fine-tuning of the implemented deep networks.
4. A comprehensive evaluation and comparison against existing state-of-the-art studies.
5. The final report will describe methodology, implementation and testing with theoretical analysis and detailed evaluation results.

Suggested Extensions

• Implementing deep hybrid networks with new architectures.
• Implementing optimal hyper-parameter selection using an automatic process.
• Theoretical analysis of the algorithms.

Reading

• Goodfellow, I., Bengio, Y. and Courville, A., 2016. Deep learning. MIT press.
• Chollet, F., 2021. Deep learning with Python.
• Journal/conference papers.

Early Deliverables

1. A report describing techniques/methods to convert a binary into an image
2. A preliminary tool to convert binaries into (grayscale) images
3. Creation of a reasonably large dataset composed of different types of malware binaries
4. A preliminary set of tests with neural networks (CNN, DNN) to classify a subset of the dataset based on their image representation

Final Deliverables

1. Leveraging neural networks for classifying malware into families
2. Describe in detail the design/architecture of the system
3. Test the system on the entire dataset and discuss the results (efficacy, efficiency) and pros/cons of each technique
4. Suggestions for improving the proposed system

Reading

• https://dl.acm.org/doi/pdf/10.1145/2016904.2016908
• https://www.intel.com/content/www/us/en/artificial-intelligence/documents/stamina-deep-learning-for-malware-protection-whitepaper.html
• https://arxiv.org/pdf/1812.07606.pdf
• https://www.blackhat.com/docs/us-15/materials/us-15-Davis-Deep-Learning-On-Disassembly.pdf
• https://arxiv.org/pdf/1903.11551.pdf
• https://www.covert.io/research-papers/deep-learning-security/Convolutional%20Neural%20Networks%20for%20Malware%20Classification.pdf
• https://dl.acm.org/doi/pdf/10.1145/2046684.2046689

Early Deliverables

1. A report describing some hard problems on lattices and how they relate
2. A report motivating the use of lattice-based cryptography
3. Proof of concept implementation of Regev's LWE-based PKE (perhaps using mathematical software e.g. Sage)

Final Deliverables

1. Implementation in C++ of Lindner-Peikert IND-CPA secure PKE based on LWE, Ring-LWE, or Module-LWE (instead of C++, also a different programming language can be used if approved by the project supervisor)
2. Use IND-CPA secure PKE as a building block to implement IND-CCA secure KEM
3. Report discussing LWE variants and justification of choice of underlying hardness assumption
4. Report explaining relevant cryptographic concepts e.g. notions of security, definitions of PKE and KEM and related primitives

Reading

• https://csrc.nist.gov/Projects/post-quantum-cryptography/Post-Quantum-Cryptography-Standardization
• https://cims.nyu.edu/~regev/papers/lwesurvey.pdf
• http://portal.acm.org/citation.cfm?id=1060590.1060603
• https://www.maths.ox.ac.uk/system/files/attachments/sage-introduction.pdf
• https://web.eecs.umich.edu/~cpeikert/pubs/lwe-analysis.pdf
• https://repo.zenk-security.com/Cryptographie%20.%20Algorithmes%20.%20Steganographie/Introduction%20to%20Modern%20Cryptography.pdf

Early Deliverables

1. A report describing provable security and how it affects parameter choices.
2. A report motivating the use of provably secure parameters.
3. Proof of concept implementation of parameterizable RSA key generation.

Final Deliverables

1. An implementation of PKCS#1 v1.5 signatures scheme with provably secure and standard parameters.
2. A benchmarking program to compare the scheme with provably secure parameters and standard ones.
3. Report explaining the implementation of RSA based schemes.
4. Report explaining relevant cryptographic concepts e.g. definitions of digital signatures, notions of security, etc.

Reading

• https://eprint.iacr.org/2018/855.pdf
• https://datatracker.ietf.org/doc/html/rfc8017
• https://www.ecrypt.eu.org/csa/documents/D5.4-FinalAlgKeySizeProt.pdf
• https://www.openssl.org/docs/manmaster/man1/openssl-speed.html

Early Deliverables

1. A report describing the state of the art in insider threat detection.
2. A clear workplan describing the set of tests to be performed, tools to be implemented and classes of techniques to be proposed and studied.

Final Deliverables

1. Design and implement a framework for development of new and existing insider threat detection techniques
2. Describe in detail the design of tests for the framework
3. Describe and analyse the features that have been used during the project to detect/stop insider threats
4. Critically compare the findings with related works

Reading

• The CERT insider threat database: https://insights.sei.cmu.edu/insider-threat/2011/08/the-cert-insider-threat-database.html
• The CERT insider threat dataset: https://resources.sei.cmu.edu/library/asset-view.cfm?assetid=508099
• Al Tabash and Happa. Insider-threat detection using gaussian mixture models and sensitivity profiles.
• Legg et al. Towards a conceptual model and reasoning structure for insider threat detection.
• Agrafiotis et al. Identifying attack patterns for insider threat detection.
• Rashid et al. A new take on detecting insider threats: exploring the use of hidden markov models.
• Legg et al. Caught in the act of an insider attack: detection and assessment of insider threat.

Early Deliverables

1. Report 1: bootstrap sampling.
2. Report 2: overview of chosen case study.
3. Report 3: prediction regions, JPRs and bootstrap JPRs.
4. Report 4: time-series models.
5. Report 5: preliminary plan, including an overview of the planned solution method and experimental evaluation plan.
6. Implement and evaluate predictor and prediction standard errors.

Final Deliverables

1. Implement code for bootstrap JPRs and evaluate empirical coverage and family-wise error rate for different hyper-parameters (e.g., k, horizon length, sample size, number of bootstrap samples).
2. Final dissertation describing problem, methods and algorithm, implementation details, computational experiments and their analysis, and conclusions.

Suggested Extensions

• 1. Compare different predictors.
• 2. Compare bootstrap JPRs with other methods mentioned in the paper (e.g., joint marginals and Scheffé)
• 3. Compare bootstrap JPRs with Monte Carlo sampling based on Bayesian neural networks and uncertainty recalibration

Reading

• (bootstrap) Chapter 6, Lecture notes of 36-402, CMU Undergraduate Advanced Data Analysis, available at https://www.stat.cmu.edu/~cshalizi/ADAfaEPoV/ADAfaEPoV.pdf
• (bootstrap JPRs) Wolf, Michael, and Dan Wunderli. "Bootstrap joint prediction regions." Journal of Time Series Analysis 36.3 (2015): 352-376.
• (time-series models) Stock Market Predictions with LSTM in Python, available at https://www.datacamp.com/community/tutorials/lstm-python-stock-market
• (time-series models) Zhang, A., Lipton, Z. C., Li, M., & Smola, A. J. (2019). Dive into deep learning, chapters 8 and 9, available at https://d2l.ai/index.html
• (extension 3) Gal Y. & Ghahramani Z., 2016, Dropout as a Bayesian Approximation: Representing Model Uncertainty in Deep Learning, ICML 2016, available at http://proceedings.mlr.press/v48/gal16.pdf
• (extension 3) Kuleshov, Volodymyr, Nathan Fenner, and Stefano Ermon. "Accurate Uncertainties for Deep Learning Using Calibrated Regression." ICML 2018, available at https://arxiv.org/pdf/1807.00263.pdf

Early Deliverables

1. Report 1: bootstrap sampling.
2. Report 2: overview of chosen case study.
3. Report 3: prediction regions, JPRs and bootstrap JPRs.
4. Report 4: time-series models.
5. Report 5: preliminary plan, including an overview of the planned solution method and experimental evaluation plan.
6. Implement and evaluate predictor and prediction standard errors.

Final Deliverables

1. Implement code for bootstrap JPRs and evaluate empirical coverage and family-wise error rate for different hyper-parameters (e.g., k, horizon length, sample size, number of bootstrap samples).
2. Final dissertation describing problem, methods and algorithm, implementation details, computational experiments and their analysis, and conclusions.

Suggested Extensions

• 1. Compare different predictors.
• 2. Compare bootstrap JPRs with other methods mentioned in the paper (e.g., joint marginals and Scheffé)
• 3. Compare bootstrap JPRs with Monte Carlo sampling based on Bayesian neural networks and uncertainty recalibration

Reading

• (bootstrap) Chapter 6, Lecture notes of 36-402, CMU Undergraduate Advanced Data Analysis, available at https://www.stat.cmu.edu/~cshalizi/ADAfaEPoV/ADAfaEPoV.pdf
• (bootstrap JPRs) Wolf, Michael, and Dan Wunderli. "Bootstrap joint prediction regions." Journal of Time Series Analysis 36.3 (2015): 352-376.
• (time-series models) Stock Market Predictions with LSTM in Python, available at https://www.datacamp.com/community/tutorials/lstm-python-stock-market
• (time-series models) Zhang, A., Lipton, Z. C., Li, M., & Smola, A. J. (2019). Dive into deep learning, chapters 8 and 9, available at https://d2l.ai/index.html
• (extension 3) Gal Y. & Ghahramani Z., 2016, Dropout as a Bayesian Approximation: Representing Model Uncertainty in Deep Learning, ICML 2016, available at http://proceedings.mlr.press/v48/gal16.pdf
• (extension 3) Kuleshov, Volodymyr, Nathan Fenner, and Stefano Ermon. "Accurate Uncertainties for Deep Learning Using Calibrated Regression." ICML 2018, available at https://arxiv.org/pdf/1807.00263.pdf

Early Deliverables

1. Report on the comparative study of the distributed platforms for processing large-scale graphs.
2. Implementation and evaluation of the proof-of-concept programs.
3. Detailed plan for the final project outcomes as agreed with the advisor.

Final Deliverables

1. The software you built to achieve the agreed project goals. The dissertation will report on different algorithms for the advanced problem you chose comparing their strengths and weaknesses. It will discuss techniques and findings, and provide meaningful visualisation of the results.

Reading

Early Deliverables

1. Report 1: overview of the artificial pancreas and underlying control algorithms.
2. Report: review of main models for glucose prediction.
3. Report: preliminary plan, including an overview of the planned solution method and experimental evaluation plan.
4. Implement code for data processing and for learning neural network-based glucose predictors. Perform training for at least one NN architecture and at least one prediction horizon.

Final Deliverables

1. Final dissertation describing problem, data, solution method, machine learning methods used, implementation details, computational experiments and their analysis, and conclusions.

Suggested Extensions

• Analysis of prediction reliability using techniques like inductive conformal prediction or approximate Bayesian inference via Monte Carlo dropout.
• Derive optimal insulin control policies using the learned glucose prediction models
• Compare the learned deep models with more classical autoregressive models.

Reading

• (for report 1) Boughton, Charlotte K., and Roman Hovorka. "Advances in artificial pancreas systems." Science translational medicine 11.484 (2019). PDF available from supervisor
• (for reports 1 and 2) Lunze, Katrin, et al. "Blood glucose control algorithms for type 1 diabetic patients: A methodological review." Biomedical signal processing and control 8.2 (2013): 107-119.
• (for report 2 and for ideas on neural net architecture) Li, Kezhi, et al. "Convolutional recurrent neural networks for glucose prediction." IEEE Journal of Biomedical and Health Informatics (2019).
• (for report 2 and for example of autoregressive model) Sparacino, Giovanni, et al. "Glucose concentration can be predicted ahead in time from continuous glucose monitoring sensor time-series." IEEE Transactions on biomedical engineering 54.5 (2007): 931-937.
• (for monte carlo dropout) Gal, Yarin, and Zoubin Ghahramani. "Dropout as a bayesian approximation: Representing model uncertainty in deep learning." international conference on machine learning. 2016.
• (for inductive conformal prediction) Papadopoulos, Harris. "Inductive conformal prediction: Theory and application to neural networks." Tools in artificial intelligence. IntechOpen, 2008.

Early Deliverables

1. Report 1: overview of the artificial pancreas and underlying control algorithms.
2. Report: review of main models for glucose prediction.
3. Report: preliminary plan, including an overview of the planned solution method and experimental evaluation plan.
4. Implement code for data processing and for learning neural network-based glucose predictors. Perform training for at least one NN architecture and at least one prediction horizon.

Final Deliverables

1. Final dissertation describing problem, data, solution method, machine learning methods used, implementation details, computational experiments and their analysis, and conclusions.

Suggested Extensions

• Analysis of prediction reliability using techniques like inductive conformal prediction or approximate Bayesian inference via Monte Carlo dropout.
• Derive optimal insulin control policies using the learned glucose prediction models
• Compare the learned deep models with more classical autoregressive models.

Reading

• (for report 1) Boughton, Charlotte K., and Roman Hovorka. "Advances in artificial pancreas systems." Science translational medicine 11.484 (2019). PDF available from supervisor
• (for reports 1 and 2) Lunze, Katrin, et al. "Blood glucose control algorithms for type 1 diabetic patients: A methodological review." Biomedical signal processing and control 8.2 (2013): 107-119.
• (for report 2 and for ideas on neural net architecture) Li, Kezhi, et al. "Convolutional recurrent neural networks for glucose prediction." IEEE Journal of Biomedical and Health Informatics (2019).
• (for report 2 and for example of autoregressive model) Sparacino, Giovanni, et al. "Glucose concentration can be predicted ahead in time from continuous glucose monitoring sensor time-series." IEEE Transactions on biomedical engineering 54.5 (2007): 931-937.
• (for monte carlo dropout) Gal, Yarin, and Zoubin Ghahramani. "Dropout as a bayesian approximation: Representing model uncertainty in deep learning." international conference on machine learning. 2016.
• (for inductive conformal prediction) Papadopoulos, Harris. "Inductive conformal prediction: Theory and application to neural networks." Tools in artificial intelligence. IntechOpen, 2008.

Early Deliverables

1. Write up of regular expressions, Thompson's construction, subset construction, minimisation.
2. Implementation design for the lexer generator, with an analysis of the data structures needed.
3. User interface design, data structure design.
4. Proof of concept implementation using hand constructed automata.

Final Deliverables

1. The project report will contain a description of the construction of automata from regular expressions and their combination into a tokeniser, written at a level suitable for a beginning final year student to understand.
2. An implementation of a lexical analyser generator which can take as input a set of regular expressions, generate a corresponding set of finite state automata, and use these to tokenise an input string.

Early Deliverables

1. A set of small worked manual examples of non-stochastic L-systems.
2. A Java program which interprets strings and draws sequences of line segments.
3. Write a report surveying classes of L-systems.

Final Deliverables

1. Demonstrate a working GUI-based L-system environment.
2. Extend your system to support stochastic productions and show how they can mimic plant growth.
3. A final report on the project activities, fully acknowledging and citing sources.

Suggested Extensions

• Investigate the relationship between L-systems and Semi-Thue systems.
• Extend your system to draw 3-dimensional objects.

Reading

• The book 'The algorithmic beauty of plants' available free from http://algorithmicbotany.org/papers/abop/abop.pdf .
• Many online resources: start with Wikipedia

Early Deliverables

1. Proof of concept program: identify one machine learning task suitable for the smart meter data and implement one learning algorithm for the chosen task.
2. Report: An overview of the datasets used. Statistics and numbers involved.
3. Report: A report on machine learning tasks for smart meter data.

Final Deliverables

1. Survey of machine learning techniques used for smart meter data analysis.
2. The developed machine learning software for smart meter data.
3. The software engineering process involved in generating your software.
4. Experiments with different smart meter data.

Suggested Extensions

• Extend the learning algorithms for online setting.
• Provide predictions with confidence measure.

Reading

• Smart*: An Open Data Set and Tools for Enabling Research in Sustainable Homes. Sean Barker, Aditya Mishra, David Irwin, Emmanuel Cecchet, Prashant Shenoy, and Jeannie Albrecht. Proceedings of the 2012 Workshop on Data Mining Applications in Sustainability (SustKDD 2012), Beijing, China, August 2012.
• Smart* Data Set for Sustainability http://traces.cs.umass.edu/index.php/Smart/Smart
• SmartMeter Energy Consumption Data in London Households https://data.london.gov.uk/dataset/smartmeter-energy-use-data-in-london-households
• UK Domestic Appliance-Level Electricity (UK-DALE) dataset http://jack-kelly.com/data/
• Revealing Household Characteristics from Smart Meter Data. Christian Beckel, Leyna Sadamori, Thorsten Staake, Silvia Santini, Energy, Vol. 78, pp. 397-410, December 2014.
• Piotr Mirowski, Sining Chen, Tin Kam Ho, Chun-Nam Yu, Demand Forecasting in Smart Grids, Bell Labs Technical Journal, 18(4), pp.135-158, 2014.

Early Deliverables

1. Proof of concept program: identify one machine learning task suitable for the smart meter data and implement one learning algorithm for the chosen task.
2. Report: An overview of the datasets used. Statistics and numbers involved.
3. Report: A report on machine learning tasks for smart meter data.

Final Deliverables

1. Survey of machine learning techniques used for smart meter data analysis.
2. The developed machine learning software for smart meter data.
3. The software engineering process involved in generating your software.
4. Experiments with different smart meter data.

Suggested Extensions

• Extend the learning algorithms for online setting.
• Provide predictions with confidence measure.

Reading

• Smart*: An Open Data Set and Tools for Enabling Research in Sustainable Homes. Sean Barker, Aditya Mishra, David Irwin, Emmanuel Cecchet, Prashant Shenoy, and Jeannie Albrecht. Proceedings of the 2012 Workshop on Data Mining Applications in Sustainability (SustKDD 2012), Beijing, China, August 2012.
• Smart* Data Set for Sustainability http://traces.cs.umass.edu/index.php/Smart/Smart
• SmartMeter Energy Consumption Data in London Households https://data.london.gov.uk/dataset/smartmeter-energy-use-data-in-london-households
• UK Domestic Appliance-Level Electricity (UK-DALE) dataset http://jack-kelly.com/data/
• Revealing Household Characteristics from Smart Meter Data. Christian Beckel, Leyna Sadamori, Thorsten Staake, Silvia Santini, Energy, Vol. 78, pp. 397-410, December 2014.
• Piotr Mirowski, Sining Chen, Tin Kam Ho, Chun-Nam Yu, Demand Forecasting in Smart Grids, Bell Labs Technical Journal, 18(4), pp.135-158, 2014.

Early Deliverables

1. Producing a simple proof-of-concept implementation of some classification and regression algorithms from scratch.
2. A report of the empirical results of such learning algorithms on a small synthetic dataset.

Final Deliverables

1. A detailed literature review of the machine learning application in WiFi fingerprinting.
2. Producing a detailed report of the dataset structure, including visual analysis (e.g. t-SNE, UMAP, etc.).
3. Implementing the algorithms on the entire competition dataset, and optimising their parameters.
4. Carefully analysing and comparing the performance accuracy of such learning algorithms.

Suggested Extensions

• Providing location estimations with some confidence measure.
• Extending the learning algorithms to an online setting, where the user provides continuous samples as she traverses the building.

Reading

• Bahl, P. and Padmanabhan, V. N. (2000). RADAR: An in-building RF-based user location and tracking system. In INFOCOM 2000. 19th Annual Joint Conference of the IEEE Computer and Communications Societies. Proceedings. IEEE (Vol. 2, pp. 775-784).
• Honkavirta, V., Perala, T., Ali-Loytty, S., and Piché, R. (2009). A comparative survey of WLAN location fingerprinting methods. In Positioning, Navigation and Communication, 2009. WPNC 2009. 6th Workshop on (pp. 243-251). IEEE.
• https://archive.ics.uci.edu/ml/datasets/UJIIndoorLoc

Early Deliverables

1. Producing a simple proof-of-concept implementation of some classification and regression algorithms from scratch.
2. A report of the empirical results of such learning algorithms on a small synthetic dataset.

Final Deliverables

1. A detailed literature review of the machine learning application in WiFi fingerprinting.
2. Producing a detailed report of the dataset structure, including visual analysis (e.g. t-SNE, UMAP, etc.).
3. Implementing the algorithms on the entire competition dataset, and optimising their parameters.
4. Carefully analysing and comparing the performance accuracy of such learning algorithms.

Suggested Extensions

• Providing location estimations with some confidence measure.
• Extending the learning algorithms to an online setting, where the user provides continuous samples as she traverses the building.

Reading

• Bahl, P. and Padmanabhan, V. N. (2000). RADAR: An in-building RF-based user location and tracking system. In INFOCOM 2000. 19th Annual Joint Conference of the IEEE Computer and Communications Societies. Proceedings. IEEE (Vol. 2, pp. 775-784).
• Honkavirta, V., Perala, T., Ali-Loytty, S., and Piché, R. (2009). A comparative survey of WLAN location fingerprinting methods. In Positioning, Navigation and Communication, 2009. WPNC 2009. 6th Workshop on (pp. 243-251). IEEE.
• https://archive.ics.uci.edu/ml/datasets/UJIIndoorLoc

Early Deliverables

1. Solution of Problems 1 and 2, their performance analysis, recommendation for cluster configuration and parameter settings, and result visualisation.
2. Reports: Map and Reduce algorithm description; experience with running your code on a Hadoop cluster; practical problems being encountered, and their solution; experience with performance analysis, and parameter tuning; recommendation for optimal parameter setting for each of the analysed problem.

Final Deliverables

1. Solution for Problems 3 and 4, their performance analysis, recommendation for cluster configuration and parameter settings, and visualisation of the results.
2. The report will provide a reflective overview of the MapReduce paradigm, and its Hadoop implementation.
3. The report will describe implementation of the Map and Reduce primitives for all four problems.
4. The report will discuss and motivate the choice of the data sets for each of the four problems.
5. If synthetic data sets were used, the report will outline the procedures used to generate the data sets.
6. The report will reflect on the experience of running, analysing and tuning your implementation on a Hadoop cluster.

Suggested Extensions

• Implement advanced data mining techniques. Apply them to interesting data sets.
• Talk to machine learning experts in the department to identify open massive scale analytics problems they are interested in solving, and implement and run them in the Hadoop framework. One such problem can be identifying trends in the epidemics spread using real-time data supplied by mobile devices. If you are interested to look at this problem, you should talk to Chris Watkins and Gregory Chockler.

Reading

• Jeffrey Dean and Sanjay Ghemawat. MapReduce: Simplified data processing on large clusters. In OSDI 2004, pages 137-150, 2004.
• Jeffrey Dean and Sanjay Ghemawat. MapReduce: a flexible data processing tool. Commun. ACM, 53(1):72-77, 2010.
• Jeffrey Dean. Experiences with MapReduce, an abstraction for large-scale computation. In PACT '06: Proceedings of the 15th international conference on Parallel architectures and compilation techniques, page 1, New York, NY, USA, 2006. ACM.
• Rajaraman, Lesocovec, Ulman. Mining of Massive Datasets. Cambridge University Press. Available online at http://infolab.stanford.edu/~ullman/mmds.html.
• MapReduce bibliography by A. Kamil, 2010: http://www.columbia.edu/~ak2834/mapreduce.html.
• Hadoop: http://hadoop.apache.org.

Early Deliverables

1. Report on Basics of Computational Complexity.
2. First report on Local Search.
3. Report on Tabu Search.
4. Implementation of Tabu Search for pseudo-boolean optimisation.

Final Deliverables

1. Full report on Local Search.
2. Report on Iterated Local Search.
3. Implementation of Iterated Local Search for pseudo-boolean optimisation.
4. Report on Genetic Algorithms.
5. Implementation of Genetic Algorithm for pseudo-boolean optimisation.
6. Report on Memetic Algorithms.
7. Implementation of Memetic Algorithm for pseudo-boolean optimisation.
8. Graphic User Interface for the four implementations.
9. Computational comparison of the four implementations.

Suggested Extensions

• Implementing metaheuristics for other optimisation problems.

Early Deliverables

1. A report describing evasion techniques used by malware
2. A report giving an overview of existing detection/mitigation techniques for evasion
3. A clear workplan describing the set of tests to be performed, and the classes of evasion mechanisms/techniques to be analysed

Final Deliverables

1. A thorough analysis of evasion techniques, e.g. against sandboxing, virtualised environments and debugging
2. Describe in detail the design of the tests
3. Describe and analyse the features that have been used during the project to detect/stop evasion
4. Critically compare the findings with related works
5. Suggestions for improving the security of sandboxes with respect to this problem

Reading

• A Survey On Automated Dynamic Malware Analysis Evasion and Counter-Evasion: PC, Mobile, and Web
• https://github.com/a0rtega/pafish
• https://github.com/LordNoteworthy/al-khaser
• https://github.com/joesecurity/pafishmacro

Early Deliverables

1. Project plan with tasks and milestones.
2. An overview of ontologies: what they are, basic technical concepts (RDF,RDFS and basic OWL).
3. Option 1: An overview of Description Logic that supports ontological reasoning.
4. Option 2: Gathering of requirements from a user, or justifying such requirements from the perspective of a potential user.
5. Initial experiments with Protege. And building a preliminary model for the chosen domain. (The project intends to utilise Protege but other available or developed tools for ontology engineering are certainly legitimate.

Final Deliverables

1. An ontology of the knowledge domain associated with the data set.
2. Option 2: a web front-end for the data set based on the ontology.
3. Option 1: A discussion of Description Logic with explicit examples in the ontology constructed.
4. Option 1: Investigate inferencing algorithms in Description Logic and their run-time complexity.
5. At least a partial encoding of the data set in the ontology.
6. Option 1: Exemplifying sophisticated inferencing using Protege (or other selected tools): using restrictions (quantifiers).
7. Option 2: The report will include examples of usage of the system.
8. Option 2: The report will include discuss the perceived advantages and disadvantages of using ontology frameworks for representing and querying the chosen data set

Suggested Extensions

• Option 2: Connect the ontology to other existing ontologies in related areas.
• Option 2: User-friendly interfaces for data representation and querying.
• Option 2: Investigate inferencing algorithms in Description Logic and their run-time complexity.
• Option 1: User-friendly interfaces for data representation and querying.
• Option 1: User-friendly interfaces for data representation and querying.
• Option 1: a web front-end for the data set based on the ontology.

Reading

• Option 2: Semantic Web for the Working Ontologist. Dean Allemang and Jim Hendler.
• Option 1: A Description Logic Primer. Markus Krotzsch, Frantiek Simancik, Ian Horrocks.
• Both options: http://protege.stanford.edu

Early Deliverables

1. A report describing the state of the art in network simulations.
2. A clear workplan describing the set of tests to be performed, tools to be implemented and classes of techniques to be proposed and studied.

Final Deliverables

1. Design and implement a framework for development of new and existing network simulation techniques
2. Describe in detail the design of tests for the framework
3. Describe and analyse the features that have been used during the project to simulate networks
4. Critically compare the findings with related works

Reading

• GNS3: https://www.gns3.com/
• https://www.sciencedirect.com/topics/computer-science/network-simulation

Early Deliverables

1. A skeleton of a mock application (in a field of your choice) with a backend, a RESTful API (for 3rd-party integrations and optionally frontend), and at least one frontend (web interface or mobile app).
2. A report describing the authentication and authorization mechanisms and best practices and how you plan to integrate these in your application.

Final Deliverables

1. The full application with integration of authentication and authorization mechanisms.
2. Complete report

Reading

• https://tools.ietf.org/html/rfc6749 (OAuth 2.0)
• https://tools.ietf.org/html/rfc6750 (Bearer Token Usage)
• https://tools.ietf.org/html/rfc6819 (OAuth 2.0 Threat Model and Security Considerations)
• https://tools.ietf.org/html/draft-ietf-oauth-security-topics-21 (Draft OAuth Security BCP)
• https://tools.ietf.org/html/draft-ietf-oauth-browser-based-apps-12 (Draft Best Practices for OAuth Browser-Based Apps)
• https://openid.net/specs/openid-connect-core-1_0.html (OpenID Connect Core 1.0)
• https://www.oauth.com/ (OAuth 2.0 Simplified)
• https://www.oauth.net/ (Community page for OAuth 2.0)
• Daniel Fett, Ralf Küsters, Guido Schmitz. 2016. A Comprehensive Formal Security Analysis of OAuth 2.0. In: 23rd ACM SIGSAC Conference on Computer and Communications Security (CCS 2016). https://dl.acm.org/doi/pdf/10.1145/2976749.2978385
• Daniel Fett, Ralf Küsters, Guido Schmitz. 2017. The Web SSO Standard OpenID Connect: In-Depth Formal Security Analysis and Security Guidelines. In: IEEE 30th Computer Security Foundations Symposium (CSF 2017). https://ieeexplore.ieee.org/iel7/8048777/8049639/08049720.pdf

Early Deliverables

1. Proof of concept program: tf-idf based similarity measurement
2. Proof of concept program: word-embedding based similarity measurement
3. (Optional) Proof of concept program: sentence-embedding based similarity measurement.
4. A report describing and assessing the results of applying the proof of concept programmes on different sentence pairs.

Final Deliverables

1. Fine-tune existing neural models (e.g. BERT), and evaluate whether they are better at dealing problems (i) and (ii) than other models.
2. A report summarising all the experiments and analysing strengths and weaknesses of each model

Suggested Extensions

• Constructing more training sentence pairs, for example by using Google Translate: given a sentence in English (denoted S), first translate it into another language (denote the translated sentence by S'), and then translate S' back to English (denote the result as S''); we should assume that S' and S'' should be similar sentences.

Reading

• Neural Network Methods for Natural Language Processing, by Y. Goldberg, Morgan and Claypool 2017
• Deep Learning, by I. Goodfellow, Y. Bengio and A. Courville, MIT Press 2016
• J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova, "BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding", arXiv pre-print 1810.04805

Early Deliverables

1. Proof of concept program: tf-idf based similarity measurement
2. Proof of concept program: word-embedding based similarity measurement
3. (Optional) Proof of concept program: sentence-embedding based similarity measurement.
4. A report describing and assessing the results of applying the proof of concept programmes on different sentence pairs.

Final Deliverables

1. Fine-tune existing neural models (e.g. BERT), and evaluate whether they are better at dealing problems (i) and (ii) than other models.
2. A report summarising all the experiments and analysing strengths and weaknesses of each model

Suggested Extensions

• Constructing more training sentence pairs, for example by using Google Translate: given a sentence in English (denoted S), first translate it into another language (denote the translated sentence by S'), and then translate S' back to English (denote the result as S''); we should assume that S' and S'' should be similar sentences.

Reading

• Neural Network Methods for Natural Language Processing, by Y. Goldberg, Morgan and Claypool 2017
• Deep Learning, by I. Goodfellow, Y. Bengio and A. Courville, MIT Press 2016
• J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova, "BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding", arXiv pre-print 1810.04805

Early Deliverables

1. Proof of concept program: tf-idf based similarity measurement
2. Proof of concept program: word-embedding based similarity measurement
3. (Optional) Proof of concept program: sentence-embedding based similarity measurement.
4. A report describing and assessing the results of applying the proof of concept programmes on different sentence pairs.

Final Deliverables

1. Fine-tune existing neural models (e.g. BERT), and evaluate whether they are better at dealing problems (i) and (ii) than other models.
2. A report summarising all the experiments and analysing strengths and weaknesses of each model

Suggested Extensions

• Constructing more training sentence pairs, for example by using Google Translate: given a sentence in English (denoted S), first translate it into another language (denote the translated sentence by S'), and then translate S' back to English (denote the result as S''); we should assume that S' and S'' should be similar sentences.

Reading

• Neural Network Methods for Natural Language Processing, by Y. Goldberg, Morgan and Claypool 2017
• Deep Learning, by I. Goodfellow, Y. Bengio and A. Courville, MIT Press 2016
• J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova, "BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding", arXiv pre-print 1810.04805

Early Deliverables

1. Proof of concept program: Long Short Term Memory (LSTM) Neural Network for text encoding.
2. (Optional) Proof of concept program: Attention Mechanism on top of (LSTM).
3. A report describing and assessing the results of applying the proof of concept programmes on data.

Final Deliverables

1. Implement an LSTM-based Encoder-Decoder network and train it with a few hundred thousand of real data
2. Augment the Encoder-Decoder network with Pointer Mechanism and Attention Mechanism, compare their performance against vanilla Encoder-Decoder
3. (Optional) Implement Transformer-based Encoder-Decoder, compare its performance against the other models
4. A report summarising all the experiments and analysing strengths and weaknesses of each model

Suggested Extensions

• Applying Reinforcement Learning (RL) on top of the Encoder-Decoder models. Recent research suggests that applying RL further boost the performance.
• Using Transfer Learning techniques to adjust existing language models, e.g. GPT-2, to generate summaries.

Reading

• Neural Network Methods for Natural Language Processing, by Y. Goldberg, Morgan and Claypool 2017
• Deep Learning, by I. Goodfellow, Y. Bengio and A. Courville, MIT Press 2016
• A. See, P. J. Liu and C. Manning, "Get To The Point: Summarization with Pointer-Generator Networks", Proceedings of the 55th ACL Annual Conference

Early Deliverables

1. Proof of concept program: Long Short Term Memory (LSTM) Neural Network for text encoding.
2. (Optional) Proof of concept program: Attention Mechanism on top of (LSTM).
3. A report describing and assessing the results of applying the proof of concept programmes on data.

Final Deliverables

1. Implement an LSTM-based Encoder-Decoder network and train it with a few hundred thousand of real data
2. Augment the Encoder-Decoder network with Pointer Mechanism and Attention Mechanism, compare their performance against vanilla Encoder-Decoder
3. (Optional) Implement Transformer-based Encoder-Decoder, compare its performance against the other models
4. A report summarising all the experiments and analysing strengths and weaknesses of each model

Suggested Extensions

• Applying Reinforcement Learning (RL) on top of the Encoder-Decoder models. Recent research suggests that applying RL further boost the performance.
• Using Transfer Learning techniques to adjust existing language models, e.g. GPT-2, to generate summaries.

Reading

• Neural Network Methods for Natural Language Processing, by Y. Goldberg, Morgan and Claypool 2017
• Deep Learning, by I. Goodfellow, Y. Bengio and A. Courville, MIT Press 2016
• A. See, P. J. Liu and C. Manning, "Get To The Point: Summarization with Pointer-Generator Networks", Proceedings of the 55th ACL Annual Conference

Early Deliverables

1. Proof of concept program: Long Short Term Memory (LSTM) Neural Network for text encoding.
2. (Optional) Proof of concept program: Attention Mechanism on top of (LSTM).
3. A report describing and assessing the results of applying the proof of concept programmes on data.

Final Deliverables

1. Implement an LSTM-based Encoder-Decoder network and train it with a few hundred thousand of real data
2. Augment the Encoder-Decoder network with Pointer Mechanism and Attention Mechanism, compare their performance against vanilla Encoder-Decoder
3. (Optional) Implement Transformer-based Encoder-Decoder, compare its performance against the other models
4. A report summarising all the experiments and analysing strengths and weaknesses of each model

Suggested Extensions

• Applying Reinforcement Learning (RL) on top of the Encoder-Decoder models. Recent research suggests that applying RL further boost the performance.
• Using Transfer Learning techniques to adjust existing language models, e.g. GPT-2, to generate summaries.

Reading

• Neural Network Methods for Natural Language Processing, by Y. Goldberg, Morgan and Claypool 2017
• Deep Learning, by I. Goodfellow, Y. Bengio and A. Courville, MIT Press 2016
• A. See, P. J. Liu and C. Manning, "Get To The Point: Summarization with Pointer-Generator Networks", Proceedings of the 55th ACL Annual Conference

Early Deliverables

1. Proof of concept programs: A "hello world" offline HTML5 application using Cache and Service Workers.
2. Proof of concept programs: A "todo list" application using IndexedDB or Web Storage.
3. Proof of concept programs: An application to draw shapes using the HTML5 canvas.
4. Proof of concept programs: A web page that loads and lists Open Street Map data in raw form.
5. Reports: Basic web page development in HTML5: HTML, JavaScript, CSS.
6. Reports: Advanced technologies: jQuery, HTML5 canvas.
7. Reports: Developing an offline HTML5 application: Web Storage, Indexed DB, Service Workers and the Cache API.
8. Reports: Open Street Map data representation and vector vs. image tile maps.

Final Deliverables

1. The program will work offline.
2. The program will allow the user to load and display map data.
3. The program will allow the user to zoom and move around the map.
4. Features for improving rendering performance such as database indexing and image caching.

Suggested Extensions

• Allowing the user to search for interesting features.
• Allowing the dynamic display of different kinds of data (E.g. highlight cafes on screen).
• The support of different OSM data formats.
• Downloading map data dynamically when a connection is present.

Reading

• Web Development Tutorials: http://www.w3schools.com/
• HTML5 Canvas: https://www.w3schools.com/graphics/canvas_intro.asp
• Service Workers: https://developer.mozilla.org/en-US/docs/Web/API/Service_Worker_API/Using_Service_Workers
• Cache API: https://developer.mozilla.org/en-US/docs/Web/API/Cache
• IndexedDB: https://developer.mozilla.org/en-US/docs/Web/API/IndexedDB_API
• Web Storage: https://developer.mozilla.org/en-US/docs/Web/API/Web_Storage_API/Using_the_Web_Storage_API
• Open Street Map: http://wiki.openstreetmap.org/wiki/Develop

Early Deliverables

1. Project plan with tasks and milestones.
2. An overview of the mechanisms that support ontological reasoning - description logic.
3. Gathering of requirements from a real user.
4. Initial experiments with an ontology language.

Final Deliverables

1. An ontology of the knowledge domain associated with the data set.
2. At least a partial encoding of the data set in the ontology.
3. A web front-end for the data set based on the ontology.
4. The report will include a justification for the chosen ontology representation based on the requirements of the user.
5. The report will include examples of usage of the system.
6. The report will include discuss the perceived advantages and disadvantages of using ontology frameworks for representing and querying the chosen data set.

Suggested Extensions

• Connect the ontology to other existing ontologies in related areas.
• User-friendly interfaces for data representation and querying.

Reading

• http://protege.stanford.edu
• http://en.wikipedia.org/wiki/Ontology_(information_science)

Early Deliverables

1. Be familiar with the machine learning algorithm you will implement. Proof of Concept (PoC) program: implement the algorithm using CPU. This PoC will help you identify the comptuations that can be parallelized in GPU.
2. Learn GPU programming with Python and CUDA. Proof of Concept programs: (i) compute the elementwise sum of two huge matrices; (ii) compute the product of two huge matrices; (iii) compute the distance between every pair of observations in a huge dataset.

Final Deliverables

1. Implement the machine learning algorithm using GPU.
2. A report on comparing the performance of the GPU program against the PoC CPU program.

Suggested Extensions

• If you choose to implement stochastic descent algorithm, there are different schemes on how coordinates or observations are sampled. Implement these schemes and compare them.
• If you choose to implement K-Means algorithm, consider distance measures other than Euclidean, e.g. Minkowski norm. Then for each cluster, a generalized centroid should be computed instead of the classical centroid. One method to compute the generalized centroid is gradient descent, which can be parallelized using GPU.

Reading

• Wes Armour. Course on CUDA Programming on NVIDIA GPUs. Online, first available in 2019.
• Yurii Nesterov. Introductory Lectures on Convex Optimization: A Basic Course. Kluwer Academic Publishers 2004.
• Yun Kuen Cheung, Richard Cole and Yixin Tao. Fully Asynchronous Stochastic Coordinate Descent: A Tight Lower Bound on the Parallelism Achieving Linear Speedup. Mathematical Programming Series A 2020.

Early Deliverables

1. Be familiar with the machine learning algorithm you will implement. Proof of Concept (PoC) program: implement the algorithm using CPU. This PoC will help you identify the comptuations that can be parallelized in GPU.
2. Learn GPU programming with Python and CUDA. Proof of Concept programs: (i) compute the elementwise sum of two huge matrices; (ii) compute the product of two huge matrices; (iii) compute the distance between every pair of observations in a huge dataset.

Final Deliverables

1. Implement the machine learning algorithm using GPU.
2. A report on comparing the performance of the GPU program against the PoC CPU program.

Suggested Extensions

• If you choose to implement stochastic descent algorithm, there are different schemes on how coordinates or observations are sampled. Implement these schemes and compare them.
• If you choose to implement K-Means algorithm, consider distance measures other than Euclidean, e.g. Minkowski norm. Then for each cluster, a generalized centroid should be computed instead of the classical centroid. One method to compute the generalized centroid is gradient descent, which can be parallelized using GPU.

Reading

• Wes Armour. Course on CUDA Programming on NVIDIA GPUs. Online, first available in 2019.
• Yurii Nesterov. Introductory Lectures on Convex Optimization: A Basic Course. Kluwer Academic Publishers 2004.
• Yun Kuen Cheung, Richard Cole and Yixin Tao. Fully Asynchronous Stochastic Coordinate Descent: A Tight Lower Bound on the Parallelism Achieving Linear Speedup. Mathematical Programming Series A 2020.

Early Deliverables

1. Code: A simple implementation of a test prioritisation algorithm
2. Code: Modifying an existing testing framework to expose code metrics for scoring paths
3. Report: Effectiveness of software testing metrics, e.g., code coverage
4. Report: Basic mechanics of symbolic execution
5. Report: Search strategies in symbolic execution

Final Deliverables

1. Code: A full test prioritisation algorithm for scoring paths based on code metrics
2. Code: At least two different strategies for paths scoring, e.g., code coverage or test execution time
3. Report: Implemented algorithms and search strategies
4. Report: Experimental evaluation and comparison of search strategies
5. Report: Responsible disclosure of vulnerabilities
6. Experiments with all search strategies on real world software. Building test harnesses to exercise multiple entry points.

Suggested Extensions

• Implement and evaluate search strategies based on genetic algorithms
• Implement and evaluate scoring metrics based on estimated bug probabilities

Reading

• Patrice Godefroid, Nils Klarlund, Koushik Sen: DART: directed automated random testing. PLDI 2005: 213-223
• Cristian Cadar, Daniel Dunbar, Dawson R. Engler: KLEE: Unassisted and Automatic Generation of High-Coverage Tests for Complex Systems Programs. OSDI 2008: 209-224

Early Deliverables

1. Proof of concept programs: Simple colourful GUI with an active button
2. Proof of concept programs: Relevant Data structures including a priority queue.
3. Proof of concept programs: Solving the eight queens problem using backtracking
4. Proof of concept program which solves a simple example of your puzzle.
5. Reports: Design Patterns for AI and Search.
6. Report on Backtracking and Recursion.
7. Reports: Constraint Satisfaction, particularly consistency techniques.
8. Reports: Techniques and Algorithms used by human solvers.
9. Reports: User Interface design for a solver
10. Reports: Complexity, NP hardness and the big O notation. Estimating problem size and hardness of your puzzle.

Final Deliverables

1. The program must have a full object-oriented design, with a full implementation life cycle using modern software engineering principles
2. The program will have a splash screen and at least two other user interaction screens.
3. The program will have a Graphical User Interface that can be used to generate, load, save, solve and help with puzzle solving/game playing.
4. The program must be able to produce puzzles that have only one correct solution, or games that have a single best strategy.
5. The report will describe the software engineering process involved in generating your software.
6. The report will describe interesting algorithms and programming techniques used (or able to be used) on the project.
7. The report will discuss the choice of data structures and their performance impact on the solving and generation algorithms.
8. The report will describe at least a backtracking and recursion solving algorithm and will compare them (including benchmark).
9. The report will describe the CSP and algorithms useful for solving the CSP including generalised arc consistency, Forward Checking, DVO and Backjumping.

Suggested Extensions

• Porting the program to a mobile device
• Solving and benchmarking public domain sets of puzzles
• Using XML based files to store and load puzzles
• Solving more than one kind of puzzle
• Using machine vision to enter puzzle
• Automatically playing the game/solving the puzzle on a public website
• The program gives hints for users when they are stuck, and allows users to step back.
• The program supports printing, different screen size, timers etc.,.

Reading

• http://ktiml.mff.cuni.cz/~bartak/constraints/

Early Deliverables

1. Precise specification of the project goals as agreed with the advisor along with data collection and analysis tools, and methodologies you plan to use to achieve the specified goals.
2. Intermediate report and programs showing and reflecting upon initial findings.

Final Deliverables

1. The software you built to achieve the project goals.
2. Final report will discuss techniques and findings, and provide meaningful visualisation of the results.

Reading

• Brendan O’Connor, Ramnath Balasubramanyan, Bryan R. Routledge, Noah A. Smith, "From Tweets to Polls: Linking Text Sentiment to Public Opinion Time Series" In Proceedings of the International AAAI Conference on Weblogs and Social Media (2010)

Early Deliverables

1. Proof of concept program: An implementation of RSA using standard data types in Java or C++ encrypting and decrypting numbers.
2. Proof of concept program: A simple program generating keys.
3. Report: A description of the RSA encryption/decryption procedure.
4. Report: Examples worked out on paper and checked using the proof of concept programme.

Final Deliverables

1. The program must have a full object-oriented design, with a full implementation life cycle using modern software engineering principles.
2. The program will use integers of arbitrary size and encrypt and decrypt test messages and files.
3. An application with a GUI will perform a task such as a secure chat or a secure file transfer over the network.
4. A combination of RSA and a symmetric encryption scheme will be implemented, where RSA will distribute the keys for the symmetric scheme.
5. The report will contain an overview of modern cryptography.
6. The report will describe RSA and number theory issues such as primality checking and factorisation.
7. The report will describe the implementation issues (such as the choice of data structures, numerical methods etc) necessary to apply the theory.
8. The report will describe the software engineering process involved in generating your software.
9. The report will contain the data on the running time of the programs.

Suggested Extensions

• An advanced symmetric encryption scheme such as DES or AES will be implemented.
• Different primality testing and factorisation methods will be implemented. Their performance will be compared and test results reported.
• Attacks on RSA and/or the symmetric scheme will be implemented and tested.

Early Deliverables

1. A report describing existing machine-learning applications of homomorphic encryption
2. The implementation of a homomorphic evaluation of a simple function e.g. weighted average on a small dataset

Final Deliverables

1. Implement the homomorphic evaluation of a function relevant to machine learning for secure genomics, inspired by recent iDash competition tasks
2. Justify the scheme / library choice and steps taken to optimise performance
3. Describe homomorphic encryption scheme, and machine learning task, underlying the solution

Reading

• https://homomorphicencryption.org
• http://www.humangenomeprivacy.org/2019/
• https://simons.berkeley.edu/talks/idash-competition
• https://github.com/Microsoft/SEAL

Early Deliverables

1. Proof of concept program: Page rank applied to a small (possibly synthetic) dataset. Its performance and efficiency should be analysed.
2. Report: An overview of the problem of document ranking and main approaches to it. An overview of PageRank.

Final Deliverables

1. Two or more different machine learning technique should be applied to a standard benchmark dataset and their performance compared.
2. The performance of the program should be optimised to enable them to process large amounts of data.
3. The report will describe the theory of underlying learning algorithms.
4. The report will describe implementation issues (such as the choice of data structures, numerical methods etc) necessary to apply the theory.
5. The report will describe computational experiments, analyse their esults and draw conclusions.

Suggested Extensions

• Comparison of pointwise, pairwise and listwise approaches.
• Use of real Internet dumps such as http://commoncrawl.org/

Reading

• Christiane Rousseau, How Google works: Markov chains and eigenvalues, URL http://blog.kleinproject.org/?p=280
• Chuan He, Cong Wang, Yi-xin Zhong, and Rui-Fan Li. A survey on learning to rank. Machine Learning and Cybernetics, 2008 International Conference on. DOI 10.1109/ICMLC.2008.4620685.
• Tie-Yan Liu, Learning to Rank for Information Retrieval, Berlin, Heidelberg: Springer. 2011

Early Deliverables

1. Proof of concept program: Page rank applied to a small (possibly synthetic) dataset. Its performance and efficiency should be analysed.
2. Report: An overview of the problem of document ranking and main approaches to it. An overview of PageRank.

Final Deliverables

1. Two or more different machine learning technique should be applied to a standard benchmark dataset and their performance compared.
2. The performance of the program should be optimised to enable them to process large amounts of data.
3. The report will describe the theory of underlying learning algorithms.
4. The report will describe implementation issues (such as the choice of data structures, numerical methods etc) necessary to apply the theory.
5. The report will describe computational experiments, analyse their esults and draw conclusions.

Suggested Extensions

• Comparison of pointwise, pairwise and listwise approaches.
• Use of real Internet dumps such as http://commoncrawl.org/

Reading

• Christiane Rousseau, How Google works: Markov chains and eigenvalues, URL http://blog.kleinproject.org/?p=280
• Chuan He, Cong Wang, Yi-xin Zhong, and Rui-Fan Li. A survey on learning to rank. Machine Learning and Cybernetics, 2008 International Conference on. DOI 10.1109/ICMLC.2008.4620685.
• Tie-Yan Liu, Learning to Rank for Information Retrieval, Berlin, Heidelberg: Springer. 2011

Early Deliverables

1. A report describing the state of the art in privacy engineering and detection of data protection violations.
2. A clear workplan describing the set of tests to be performed, tools to be implemented and classes of detection mechanisms/techniques to be proposed and studied.

Final Deliverables

1. Design and implement a framework for development of new and existing techniques to detect information sharing compliance violations
2. Describe in detail the design of tests
3. Describe and analyse the features that have been used during the project to detect/stop violations
4. Critically compare the findings with related works

Reading

• Adam Barth, Anupam Datta, John C Mitchell, and Helen Nissenbaum. 2006. Privacy and contextual integrity: Framework and applications. In Security and Privacy, 2006 IEEE Symposium on. IEEE, pp. 184–198
• David Basin, Søren Debois, and Thomas Hildebrandt. 2018. On Purpose and by Necessity: Compliance under the GDPR. Presentation.
• David Basin, Felix Klaedtke, Srdjan Marinovic, and Eugen Zalinescu. 2015. Monitoring of temporal first-order properties with aggregations. Formal methods in system design 46, 3 (2015), 262–285.
• Sean Brooks, Sean Brooks, Michael Garcia, Naomi Lefkovitz, Suzanne Lightman, and Ellen Nadeau. 2017. An Introduction to Privacy Engineering and Risk Management in Federal Systems. US Department of Commerce, National Institute of Standards and Technology.
• Ann Cavoukian et al. 2014. Privacy by design documentation for software engineers version 1.0. PbD-SE) Burlington, MA: Organization for the Advancement of Structured Information Standards (OASIS), work in progress (2014).
• Anupam Datta, Jeremiah Blocki, Nicolas Christin, Henry DeYoung, Deepak Garg, Limin Jia, Dilsun Kaynar, and Arunesh Sinha. 2011. Understanding and protecting privacy: Formal semantics and principled audit mechanisms. In International Conference on Information Systems Security. Springer, 1–27.
• European Commission. 2018. General Data Protection Regulation. https://ec.europa.eu/commission/priorities/justice-and-fundamentalrights/data-protection/2018-reform-eu-data-protection-rules_en
• Information Exchange Policy 1.0. https://www.first.org/iep/.
• Privacy Management Reference Model (PMRM) https://www.oasis-open.org/committees/pmrm/
• Gina Fisk, Calvin Ardi, Neale Pickett, John Heidemann, Mike Fisk, and Christos Papadopoulos. 2015. Privacy principles for sharing cyber security data. In Security and Privacy Workshops (SPW), 2015 IEEE. IEEE, 193–197.
• Jassim Happa, Nick Moffat, Michael Goldsmith, and Sadie Creese. 2018. Run-Time Monitoring of Data-Handling Violations. In Computer Security. Springer, 213–232.
• Bert-Jaap Koops and Ronald Leenes. 2014. Privacy regulation cannot be hardcoded. A critical comment on the 'privacy by design' provision in data-protection law. International Review of Law, Computers and Technology 28, 2 (2014), 159–171.
• Lauren B Movius and Nathalie Krup. 2009. US and EU privacy policy: comparison of regulatory approaches. International Journal of Communication 3 (2009), 19.
• Stuart Shapiro, Nickyra Washington, Kristen Miller, Julie Snyder, and Julie McEwen. 2014. MITRE Privacy Engineering Framework. MITRE Corporation Technical Report.
• Valeria Soto-Mendoza, Patricia Serrano-Alvarado, Emmanuel Desmontils, and Jose Antonio Garcia-Macias. 2015. Policies composition based on data usage context. In Sixth International Workshop on Consuming Linked Data (COLD2015) at ISWC.
• UK Information Commissioner's Office. 2016. How do we apply legitimate interests in practice? https://ico.org.uk/for-organisations/guide-to-data-protection/guide-to-the-general-data-protection-regulation-gdpr/legitimate-interests/how-do-we-apply-legitimate-interests-in-practice/.

Early Deliverables

1. Report: An overview of ridge regression describing the concepts 'training set' and 'test set', giving the formulas for a regression algorithm and defining all terms in them.
2. Proof of concept program: a regression algorithm applied to a small artificial dataset.
3. Report: Examples of applications of Ridge Regression worked out on paper and checked using the prototype program.
4. Loading a real-world dataset, simple pre-processing and visualisation.

Final Deliverables

1. The program must have a full object-oriented design, with a full implementation life cycle using modern software engineering principles.
2. The program will work with a real-world dataset, read the data from a file, preprocess the data and apply regression using parameters selected by the user with parameters entered by the user.
3. The program will have a graphical user interface.
4. The program will automatically perform tests such as comparison of different kernels, parameters, etc and visualise the results.
5. The program will implement another learning algorithm such as nearest neighbours or neural networks, and compare regression against it.
6. The report will describe the theory of regression and derive the formulas.
7. The report will describe the implementation issues (such as the choice of data structures, numerical methods etc) necessary to apply the theory.
8. The report will describe the software engineering process involved in generating your software.
9. The report will describe computational experiments with different kernels and parameters and draw conclusions.
10. The report will describe the competitor algorithm(s), compare the performance and draw conclusions.

Suggested Extensions

• Use of regression in the on-line mode with growing and sliding window.
• Application of regression to different datasets (including those found by the student).
• Implementation of regression-type algorithms such as Kernel Aggregating Algorithm Regression, Kernel Aggregating Algorithm Regression with Changing Dependencies etc and computational experiments with them.

Early Deliverables

1. Proof of concept program: Implementation of exact policy optimisation by value iteration for a general MDP.
2. Proof of concept program: Implementation of Q-learning for grid world.
3. Proof of concept program: GUI built with buttons etc.,
4. Proof of concept program: MVC program with with controls.
5. Report on Markov decision processes (MDPs)
6. Report on notions of policy and value function; optimal policy/value function via the Bellman equation. Finding optimal policies by value iteration and policy iteration
7. Report on Concept of learning as incrementally optimising policy in a MDP
8. Report on Q-learning

Final Deliverables

1. The program must have a full object-oriented design, with a full implementation life cycle using modern software engineering principles
2. The program will at least demonstrate Q-learning with various exploration strategies in a simple world.
3. The program must have a text based control file from which it can be configured.
4. The program will have a Graphical User Interface that can be used to animate simulations.
5. The report will describe the software engineering process involved in generating your software.
6. The report will include a description of the theory of MDPs, optimal policies and value functions via the Bellman equation, dynamic programming methods of finding optimal policies, and Q-learning.
7. The report will include some experiments that compare the effects of different parameter choices on the progress of Q-learning: over-optimistic values causing excessive exploration, and pessimistic initial values causing the development of superstitions.

Suggested Extensions

• Implementing RL for more a more challenging problem than a grid world, such as an acrobat.
• Experimental analysis of the performance of different learning algorithms is interesting but challenging

Reading

• Chapter on Reinforcement Learning in 'Machine Learning' by Tom Mitchell

Early Deliverables

1. Proof of concept program: Kernel Ridge Regression applied to a small artificial dataset.
2. Report: An overview of ridge regression describing the concepts 'training set' and 'test set', giving the formulas for Ridge Regression and defining all terms in them.
3. Report: Examples of applications of Ridge Regression worked out on paper and checked using the prototype program.

Final Deliverables

1. The program will work with a real dataset such as Boston housing, read the data from a file, preprocess the data (normalise, split the dataset into the test and training sets) and apply Ridge Regression using a kernel with parameters.
2. The program will automatically perform tests such as comparison of different kernels, parameters, etc. The results will be visualised.
3. The program will work in batch and on-line modes.
4. Tests will be performed to compare plain Ridge Regression against other regression-based algorithms such as Kernel Aggregating Algorithm Regression, Kernel Aggregating Algorithm Regression with Changing Dependencies, or gradient descent.
5. The report will describe the theory of Ridge Regression and derive the formulas.
6. The report will describe implementation issues (such as the choice of data structures, numerical methods etc) necessary to apply the theory.
7. The report will describe computational experiments, analyse their results and draw conclusions

Suggested Extensions

• Application of Ridge Regression to different datasets (including those suggested by the student).
• Comparison with other learning algorithms (e.g., aggregating algorithm regression, neural networks, gradient descent-based etc).
• Combining regression with methods of prediction with expert advice.

Reading

• Y.Kalnishkan, Kernel Methods, http://onlineprediction.net/?n=Main.KernelMethods
• J.Shawe-Taylor and N.Cristianini, Kernel Methods for Pattern Analysis Cambridge University Press, 2004
• C.E.Rasmussen and C.Williams, Gaussian Processes for Machine Learning, the MIT Press, 2006 http://www.gaussianprocess.org/
• B.Schlkopf and A.J.Smola, Learning with Kernels: Support Vector Machines, Regularization, Optimization, and Beyond, The MIT Press, 2001.

Early Deliverables

1. Proof of concept program: Kernel Ridge Regression applied to a small artificial dataset.
2. Report: An overview of ridge regression describing the concepts 'training set' and 'test set', giving the formulas for Ridge Regression and defining all terms in them.
3. Report: Examples of applications of Ridge Regression worked out on paper and checked using the prototype program.

Final Deliverables

1. The program will work with a real dataset such as Boston housing, read the data from a file, preprocess the data (normalise, split the dataset into the test and training sets) and apply Ridge Regression using a kernel with parameters.
2. The program will automatically perform tests such as comparison of different kernels, parameters, etc. The results will be visualised.
3. The program will work in batch and on-line modes.
4. Tests will be performed to compare plain Ridge Regression against other regression-based algorithms such as Kernel Aggregating Algorithm Regression, Kernel Aggregating Algorithm Regression with Changing Dependencies, or gradient descent.
5. The report will describe the theory of Ridge Regression and derive the formulas.
6. The report will describe implementation issues (such as the choice of data structures, numerical methods etc) necessary to apply the theory.
7. The report will describe computational experiments, analyse their results and draw conclusions

Suggested Extensions

• Application of Ridge Regression to different datasets (including those suggested by the student).
• Comparison with other learning algorithms (e.g., aggregating algorithm regression, neural networks, gradient descent-based etc).
• Combining regression with methods of prediction with expert advice.

Reading

• Y.Kalnishkan, Kernel Methods, http://onlineprediction.net/?n=Main.KernelMethods
• J.Shawe-Taylor and N.Cristianini, Kernel Methods for Pattern Analysis Cambridge University Press, 2004
• C.E.Rasmussen and C.Williams, Gaussian Processes for Machine Learning, the MIT Press, 2006 http://www.gaussianprocess.org/
• B.Schlkopf and A.J.Smola, Learning with Kernels: Support Vector Machines, Regularization, Optimization, and Beyond, The MIT Press, 2001.

Early Deliverables

1. Report: overview of SAT, SMT and SMT solving techniques.
2. Report: bounded model checking and encoding into SMT.
3. Install z3 and Python APIs. Develop code for SMT solving of a simple example problem (e.g., Sudoku, or job shop scheduling).
4. Develop code for BMC of generic transition systems.

Final Deliverables

1. Develop code for encoding a planning problem on a 2-d grid with obstacle as a BMC problem.
2. Evaluate your solution with arbitrary and randomly generated planning instances.
3. Analyze SMT solver performance for different grid sizes and numbers of obstacles.
4. Final dissertation. It should include background on SAT, SMT, BMC, and AI planning, as well as the description of your solution for the encoding of AI planning into SMT, some example solutions obtained with the developed planning method, and performance analysis results.

Suggested Extensions

• Extend the planning problem and your solution with transition costs and cost constraints.
• Develop code to visualize path solutions on the 2d grid.

Reading

• De Moura, Leonardo, and Nikolaj Bjørner. "Satisfiability modulo theories: introduction and applications." Communications of the ACM 54.9 (2011): 69-77. (available at https://dl.acm.org/doi/abs/10.1145/1995376.1995394)
• De Moura, Leonardo, and Nikolaj Bjørner. "Satisfiability modulo theories: An appetizer." Brazilian Symposium on Formal Methods (2009): 23-36. (available at https://link.springer.com/chapter/10.1007/978-3-642-10452-7_3)
• Amla, Nina, et al. "An analysis of SAT-based model checking techniques in an industrial environment." (up to section 2.3 included). (available at https://link.springer.com/chapter/10.1007/11560548_20)
• Lectures on " Introduction to planning" and " Planning formalisms" of the " Autonomous Intelligent Systems" course
• Z3 tool page, https://github.com/Z3Prover/z3
• Programming Z3 tutorial (sections 1,2,3,5), https://theory.stanford.edu/~nikolaj/programmingz3.html
• Jupyter notebooks illustrating Z3's Python APIs (includes code for Programming Z3 tutorial), https://notebooks.azure.com/nbjorner/projects/z3examples
• Online Z3 tutorial (using SMT-lib input format), https://rise4fun.com/z3/tutorial/guide
• Dennis Yurichev, "SAT/SMT by Examples", https://yurichev.com/writings/SAT_SMT_by_example.pdf (a huge collection of maths and CS problems solved with SAT or SMT)

Early Deliverables

1. Report: overview of SAT, SMT and SMT solving techniques.
2. Report: bounded model checking and encoding into SMT.
3. Install z3 and Python APIs. Develop code for SMT solving of a simple example problem (e.g., Sudoku, or job shop scheduling).
4. Develop code for BMC of generic transition systems.

Final Deliverables

1. Develop code for encoding a planning problem on a 2-d grid with obstacle as a BMC problem.
2. Evaluate your solution with arbitrary and randomly generated planning instances.
3. Analyze SMT solver performance for different grid sizes and numbers of obstacles.
4. Final dissertation. It should include background on SAT, SMT, BMC, and AI planning, as well as the description of your solution for the encoding of AI planning into SMT, some example solutions obtained with the developed planning method, and performance analysis results.

Suggested Extensions

• Extend the planning problem and your solution with transition costs and cost constraints.
• Develop code to visualize path solutions on the 2d grid.

Reading

• De Moura, Leonardo, and Nikolaj Bjørner. "Satisfiability modulo theories: introduction and applications." Communications of the ACM 54.9 (2011): 69-77. (available at https://dl.acm.org/doi/abs/10.1145/1995376.1995394)
• De Moura, Leonardo, and Nikolaj Bjørner. "Satisfiability modulo theories: An appetizer." Brazilian Symposium on Formal Methods (2009): 23-36. (available at https://link.springer.com/chapter/10.1007/978-3-642-10452-7_3)
• Amla, Nina, et al. "An analysis of SAT-based model checking techniques in an industrial environment." (up to section 2.3 included). (available at https://link.springer.com/chapter/10.1007/11560548_20)
• Lectures on " Introduction to planning" and " Planning formalisms" of the " Autonomous Intelligent Systems" course
• Z3 tool page, https://github.com/Z3Prover/z3
• Programming Z3 tutorial (sections 1,2,3,5), https://theory.stanford.edu/~nikolaj/programmingz3.html
• Jupyter notebooks illustrating Z3's Python APIs (includes code for Programming Z3 tutorial), https://notebooks.azure.com/nbjorner/projects/z3examples
• Online Z3 tutorial (using SMT-lib input format), https://rise4fun.com/z3/tutorial/guide
• Dennis Yurichev, "SAT/SMT by Examples", https://yurichev.com/writings/SAT_SMT_by_example.pdf (a huge collection of maths and CS problems solved with SAT or SMT)

Early Deliverables

1. A report giving an overview of prominent existing key value stores such as Chord and SkipGraph.
2. A report describing how existing key value stores deal with environmental changes.
3. A proof-of-concept program implementing a simple (possibly inefficient) distributed key value store.

Final Deliverables

1. A distributed key value store that provides standard dictionary operations and is scalable to large data sets.
2. A report describing the architecture of the implemented key value store.
3. A report on the performed experiments and simulation results.

Suggested Extensions

• Supporting advanced dictionary operations, e.g., range queries.
• More efficient storage of data by using coding techniques.
• Simulating node churn.

Reading

• Ion Stoica, Robert Morris, David R. Karger, M. Frans Kaashoek, Hari Balakrishnan: Chord: A scalable peer-to-peer lookup service for internet applications. SIGCOMM 2001: 149-160
• James Aspnes, Gauri Shah: Skip graphs. ACM Trans. Algorithms 3(4): 37 (2007)

Early Deliverables

1. Report on Fixed Parameter Tractability and how it is different from the P vs NP tradition, with examples.
2. About Workflow Scheduling subject to access control: Where it occurs, why it is important and the kinds of constraints and authorisations that are useful.
3. Report on the analysis of the Fixed Parameter Tractability and the NP-hardness of WSP
4. Report on various other constraint satisfaction techniques used to solve industrial problems such as SAT and CSP solvers.
5. Programs: A simple backtracking algorithm, an FPT algorithm for the Vertex Cover problem.

Final Deliverables

1. The program must have a full object-oriented design, with a full implementation life cycle using modern software engineering principles
2. The program should have an excellent GUI for experimentation.
3. The report should include experimental assessment of the FPT algorithm, possibly in comparison with a SAT or CSP solver.
4. The report will describe the software engineering process involved in generating your software.
5. The report will describe and analyse your implementation of the backtracking FPT algorithm for solving WSP with user-independent constraints.

Suggested Extensions

• Study of the Valued WSP allowing us to provide solutions even when WSP is unsatisfiable.

Early Deliverables

1. Implementation of the first-price sealed-bid auction (FPSBA) functionality in one existing MPC framework
2. Implementation of REST interfaces to submit bids to and receive results from auction servers
3. Implementation of controller node to configure the auction platform (e.g., which items should be auctioned and how long the auction runs)

Final Deliverables

1. User-friendly web interface
2. Implementation of client-side secret sharing + communication with auction servers

Reading

• Marcella Hastings, Brett Hemenway, Daniel Noble, Steve Zdancewic: "SoK: General Purpose Compilers for Secure Multi-Party Computation". In IEEE Symposium on Security and Privacy 2019.
• Marcel Keller: "MP-SPDZ: A Versatile Framework for Multi-Party Computation". In CCS 2020.
• Lennart Braun, Daniel Demmler, Thomas Schneider, Oleksandr Tkachenko: "MOTION - A Framework for Mixed-Protocol Multi-Party Computation". In ACM TOPS 2022.
• https://github.com/data61/MP-SPDZ
• https://github.com/encryptogroup/MOTION

Early Deliverables

1. A report describing current serverless offerings from Google (G Cloud), Microsoft (Azure) and Amazon (AWS)
2. A report with an overview access control systems for G Cloud, Azure and AWS
3. A proof of concept function that executes in one of these clouds

Final Deliverables

1. A detailed descriptoin of the access control system of one of the cloud providers
2. Examples of functions that represent each of the risks described
3. A detailed description of a set of policies and recommendations so developers avoid these risks
4. A critical comparison with findings from related works

Reading

• https://cloud.google.com
• https://aws.amazon.com
• https://azure.microsoft.com
• https://dl.acm.org/doi/abs/10.1145/3366423.3380173
• https://www.usenix.org/conference/hotcloud19/presentation/obetz
• https://dl.acm.org/doi/abs/10.1145/3366615.3368351>

Early Deliverables

1. Given a maze game of your interest (e.g. Pac-Man), design and implement a simple graphical visualisation for it.
2. Consider an agent inhabiting the maze world. Model the agent's navigation task as a search problem. Implement a graph search version of depth-first search and breadth-first search to reach a specific location in the maze. Run experiments with three different size maze: small, medium and large. Visualise the agent following the plan step-by-step.
3. Now change the cost function in such a way that traversing some areas is more expensive than traversing other areas (e.g. Pac-Man: you can charge more for dangerous steps in areas close to the ghosts). Implement the uniform-cost graph search algorithm to reach a specific location in the maze. Run experiments with three different size maze: small, medium and large. Visualise the agent following the plan step-by-step.

Final Deliverables

1. Implement the algorithm A* to find the shortest path through the maze that touches all four corners. Use two different, consistent heuristic functions. Run experiments with three different size maze: small, medium and large. Visualise the agent following the plan step-by-step.
2. Assume each cell in the maze has an item to collect. Formulate a search problem for collecting all items (e.g. in Pac-Man: eating all the food) and implement A* to do that in as few steps as possible. Also implement a greedy search and compare the results of the two algorithms in three different size maze: small, medium and large. Visualise the actions of the agent.
3. Now assume that you have additional agents in your maze, one is the main player and the other are the opponents (e.g. Pac-Man and the ghosts). Model this new problem and implement the player as a reflex agent that tries to visits all cells avoiding the opponents.
4. Implement an adversarial Minmax search agent for each agent; start with the player and one opponent and then increase the number of opponents. You need to write an evaluation function that evaluates states. Run experiments with three different size maze: small, medium and large. Visualise the actions of the agents.
5. Design and implement a new agent that uses Alpha-beta pruning to more efficiently explore the minimax tree. Run experiments with three different size maze: small, medium and large. Visualise the actions of the agents.
6. Design and implement an Expectimax agent to model the probabilistic behaviour of an agents who makes suboptimal choices. Run experiments with three different size maze: small, medium and large. Visualise the actions of the agents.

Suggested Extensions

• Use value iteration and Q-learning to underpin the behaviour of the player.

Reading

• Stuart Russell and Peter Norvig, AI: A Modern Approach, 3rd Edition.

Early Deliverables

1. You are to choose either Option 1 (heuristics and approximations) or Option 2 (SAT-solving).
2. Report: A full formal description of the problem.
3. Report: P vs NP, NP-completeness, NP-hardness
4. Option 1: Report: Strategies for solving NP-complete problems, such as: heuristics; approximation algorithms; tractable special cases; algorithms that are slow in the worst case but fast on some 'non-worst-case' inputs.
5. Option 1: Report: Fast in theory, slow in practice. Merge sort and heap sort are sorting algorithms with excellent theoretical worst-case guarantees, but they may find themselves beaten in practice by quicksort, which has worse worst-case guarantees. Another example: linear programming is the basis of many planning systems, and there are algorithms for solving LPs in polynomial time, but these are not used in practice; instead, algorithms are used that have worst-case exponential-time behaviour. Report and comment on this phenomenon.
6. Option 1: Proof of concept program: A program that generates a random instance of Vertex Cover, shows it to the user, reads a suggested solution from the user, and verifies whether this solution is correct.
7. Option 1: Proof of concept program: A simple greedy heuristic
8. Option 1: Proof of concept program: The simple 2-approximation algorithm
9. Option 2: Encoding and data structures: write a program that reads a CNF from a file in the DIMACS format and a truth-assignment (generated randomly), and outputs true or false depending if the assignment satisfies the input CNF or not.
10. Option 2: Investigate the DPLL algorithm and write a report that shows manually worked examples of its operation. The report should include the pseudo code of the DPLL algorithm, providing also diagrammatic examples of its run on simple example input CNF formulas. Explain the algorithm and its running time.
11. Option 2: Write procedures that will serve as building blocks for the e cental SAT solver. Specifically, write a procedure that gets as an input a CNF formula and a *partial* assignment to its variable and outputs the new CNF formula resulting from assigning to the original formula. This should include simplifying (getting rid) of satisfied clauses, if they exist, and should output False if the CNF formula becomes false under the given assignment.

Final Deliverables

1. Option 1: A program with a GUI that can show an instance on the screen and illustrate a solution for it. Your program should be able to read a problem instance from disk, in a format taken from an existing 'benchmark library', and it should also be able to generate an instance of the problem at random with specified parameters.
2. Option 1: The program should offer at least three different solution strategies, such as a heuristic, an approximation algorithm, and a backtracking search 'branching algorithm'.
3. Option 1: The report should contain a section on the outcomes of experiments done between the different algorithms for instances of some benchmark library, with conclusions drawn about which algorithms seem suitable in different situations.
4. Option 2: Write a program which solves SAT using the DPLL algorithm. It is advisable to write the code in a language like C++ or C. But it is also possible to implement it in java or python, as the program is intended for educational purposes.
5. Option 2: Demonstrate the run of the SAT-solver on substantial instances (taken from known benchmarks).
6. Option 2: Investigate the conflict-driven clause learning heuristic (CDCL). Show in your report manually worked examples of DPLL with the clause learning heuristic as well a pseudocode for CDCL.
7. Option 2: Extend your DPLL solver to uses some kind of very simple clause learning heuristic, and explain your choice, showing how it improves the run time on some of the benchmarks.
8. Final report should contain a section on the outcomes of experiments done between the different algorithms/SAT-solvers for instances of some benchmark library, with conclusions drawn about which algorithms seem suitable in different situations.

Suggested Extensions

• A more careful theory report, including notions of 'problem reductions' for NP-complete problems. For Vertex Cover, this should at least cover the two problems Independent Set and Max Clique.
• A larger range of algorithms, or algorithms using more advanced concepts. One advanced approach for this problem would be the use of an LP relaxation. Another advanced approach would be to use methods from parameterized complexity (you may ask for references here).
• Careful optimisation of one particular algorithm for the problem, for example a carefully tweaked branching algorithm, to make it as fast as possible across a range of instances.
• Investigations into 'lower bounds' for your problem, to test how far off each algorithm is from an optimal solution.
• For option 2: Extend the DPLL SAT-solver to include a fully functional conflict driven clause learning (CDCL). The full specification for this could be taken from the Handbook (see references below) or from the Wikipedia page. Apply these extensions to the SAT-solver and systematically compare the run-time improvement.
• For option 2: Extend the DPLL SAT-solver to include Restarts. The full specification for this could be taken from the Handbook (see references below) or from the Wikipedia page. Apply these extensions to the SAT-solver and systematically compare the run-time improvement.
• Investigate the notion of Satisfiability Modulo a Theory (SMT). Show how it extends SAT. Apply this to your tool for a chosen theory and systematically compare the run-time improvement on relevant instances.
• Investigate the theoretical foundation of SAT solving: the resolution refutation system and its provable limitations.

Reading

• Selected parts of Handbook of Satisfiability. Editors: Biere, A., Heule, M., Van Maaren, H., Walsh, T. Feb 2009. Volume 185 of Frontiers in Artificial Intelligence and Applications.
• https://en.wikipedia.org/wiki/Conflict-Driven_Clause_Learning

Early Deliverables

1. Report: A full formal description of the problem
2. Report: P vs NP, NP-completeness, NP-hardness
3. Report: Strategies for solving NP-complete problems, such as: heuristics; approximation algorithms; tractable special cases; algorithms that are slow in the worst case but fast on some 'non-worst-case' inputs
4. Report: Fast in theory, slow in practice. Merge sort and heap sort are sorting algorithms with excellent theoretical worst-case guarantees, but they may find themselves beaten in practice by quicksort, which has worse worst-case guarantees. Another example: linear programming is the basis of many planning systems, and there are algorithms for solving LPs in polynomial time, but these are not used in practice; instead, algorithms are used that have worst-case exponential-time behaviour. Report and comment on this phenomenon.
5. Proof of concept program: A program that takes a graph and either finds a cycle in it or reports that the graph contains no cycles.
6. Proof of concept program: A 'cleaning algorithm' to remove vertices of degree 1 and 2 and self-loops
7. Proof of concept program: A simple greedy heuristic, with or without using the cleaning step
8. Proof of concept program: An approximation algorithm, for example the 3-approximation described in the lecture notes at https://www.cs.cmu.edu/afs/cs/academic/class/15854-f05/www/scribe/lec9.pdf

Final Deliverables

1. A program with a GUI that can show an instance on the screen and illustrate a solution for it. Your program should be able to read a problem instance from disk, in a format taken from an existing 'benchmark library', and it should also be able to generate an instance of the problem at random with specified parameters.
2. The program should offer at least three different solution strategies, such as a heuristic, an approximation algorithm, and a backtracking search 'branching algorithm'.
3. The report should contain a section on the outcomes of experiments done between the different algorithms for instances of some benchmark library, with conclusions drawn about which algorithms seem suitable in different situations.

Suggested Extensions

• A more careful theory report, including notions of 'problem reductions' for NP-complete problems. For Feedback Vertex Set, this should include at least the hardness reduction from Vertex Cover.
• A larger range of algorithms, or algorithms using more advanced concepts. One advanced approach for this problem would be the use of an LP relaxation. Another advanced approach would be to use methods from parameterized complexity (you may ask for references here).
• Careful optimisation of one particular algorithm for the problem, for example a carefully tweaked branching algorithm, to make it as fast as possible across a range of instances.
• Investigations into 'lower bounds' for your problem, to test how far off each algorithm is from an optimal solution.

Reading

• A good set of benchmark instances are found on the following page: https://pacechallenge.wordpress.com/pace-2016/track-b-feedback-vertex-set/

Early Deliverables

1. You will need to work with datasets from different sources. Be familiar with the datasets. Identify the attributes which are strong indicators of good players in every position.
2. Learn the powerful visualization library ggplot2 (or plotnine if you are using Python). In particular, learn how to represent high-dimensional data using bubble plot, and/or via PCA.
3. Make a plan how you are going to use the datasets to complete a mission that can help your sports team. The proposal should indicate how the data should be pre-processed, which learning algorithms you will use, and how the results of data analysis will be used.

Final Deliverables

1. Generate simple but informative visual illuminations which allow you to present your data analysis results to your colleagues.
2. Implement a software with user-friendly interface, which allows your colleagues to explore the data themselves. The software should achieve the followings: (i) recommend which players your team can bring in, based on the constraints on position, age and budget your colleagues inputs; (ii) provide your colleagues of customizable visual illuminations, and (iii) display annotated tables of detailed data when your colleague wants to look into it.

Early Deliverables

1. Report - Protein Structures
2. Report - The Protein Data Bank and the file formats used in accessing it.
3. Proof of Concept - distributing protein structure files amongst MapReduce cluster.
4. Proof of Concept - Running MapReduce using a single type of executable for analysing protein structures.

Final Deliverables

1. The final framework will allow users to input a map and reduce step (namely an executable to run on an individual protein structure in the former and a parser for the generated log file from the reduce step) to analyse a set of protein structures.
2. The developer will provide a clear set of guidelines for the map/reduce executables.
3. The final data should be returnable in variety of different formats.

Suggested Extensions

• The PDB is updated on a regular basis. The data set can mirror the official database and is distributed amongst the MapReduce cluster.
• A set of libraries so that arbitrary executables can be easily put into the MapReduce formalism.
• The PDB is comprised of many different types of structure (DNA and RNA macro-molecules are included as well as proteins). An interface to select fractions of the PDB for analysis.
• Implementation on a Public Mapreduce cluster e.g. Amazon.

Reading

• The protein data bank H.M.Berman et al. Nucl. Acid. Res. (2000) 28(1): 235-242. Doi: 10.1093/nar/28.1.235

Early Deliverables

1. Code: Instrumentation for statement and function coverage and basic visualisation
2. Report: Java Bytecode and Soot's intermediate representation
3. Report: Code instrumentation
4. Report: Survey on coverage metrics

Final Deliverables

1. Code: Branch coverage, condition coverage, and MC/DC coverage. GUI for running and visualising the analysis.
2. Report: Architecture of Soot
3. Report: Design and implementation of the instrumentation tool
4. Report: Test requirements for safety critical systems
5. Formal description and practical implementation of a custom coverage metric.

Suggested Extensions

• Computing cyclomatic complexity using Soot's analysis components

Reading

• Lam et al: The Soot framework for Java program analysis: a retrospective. Cetus Users and Compiler Infastructure Workshop (CETUS 2011)
• The Soot Wiki: https://github.com/Sable/soot/wiki

Early Deliverables

1. A report on task allocation algorithms in current usage and the degree to which trust is implicit in their design
2. A report on the current state of the art in anomaly detection (behavioural and statistical) in task allocation systems. This may extend to underlying network infrastructure (for example selfish routing attacks in Mobile Ad hoc Networks (MANETs) performed by malicious nodes in an attempt to gain intelligence about the larger mission)
3. A clear workplan describing the set of tests to be performed, and a specific use case defining what type of network (MANET, traditional wireless/wired, Internet-based) and task allocation protocol will form the focus of experiments

Final Deliverables

1. A simulation or experimental demonstration of the principles explored in the project, such as a simple (half dozen nodes) decision making system with attack and mitigation demonstrations
2. A final report documenting the research, design, and implementation of the system. This report will extend to cover the experimental design, and how the choices made in the systems design phase informed experimental choices (and vice versa). A clear grasp of experimental process and hypothesis testing should be demonstrated through appropriate use of tools and techniques

Reading

• https://ieeexplore.ieee.org/abstract/document/9002704
• https://ieeexplore.ieee.org/abstract/document/7776949
• https://gala.gre.ac.uk/id/eprint/16178/
• https://ieeexplore.ieee.org/abstract/document/8758417
• https://ieeexplore.ieee.org/abstract/document/7809102

Early Deliverables

1. A report describing architectural components of a PKI system and addressing Blockchain technologies potentials
2. A design of the PKI data model
3. A proof of concept implementation of the auditing system using a DLT technologies
4. Testing the Blockchain-based PKI system

Final Deliverables

1. Design and develop the data model of PKI
2. Develop a prototype of PKI system using Blockchain
3. Complete report
4. The report will have a User Manual

Reading

• https://medium.com/swlh/top-5-dlt-blockchain-protocols-to-consider-in-your-project-dae7fc6d381d
• https://www.ibm.com/blogs/research/2018/02/architecture-hyperledger-fabric/
• https://hyperledger-fabric.readthedocs.io/en/release-1.4/getting_started.html
• https://medium.com/swlh/top-5-dlt-blockchain-protocols-to-consider-in-your-project-dae7fc6d381d
• https://pki-tutorial.readthedocs.io/en/latest/simple/
• www.openssl.org
• EJBCA https://www.ejbca.org
• Dierks, T., (2008). The transport layer security (TLS) protocol version 1.2, RFC-5246
• Vacca, J. R. (2004). Public Key Infrastructure. Boca Raton: C R C Press LLC.
• Cooper, D., Santesson, S., Farrell, S., Boeyen, S., Housley, R. Polk, W. (2008) Internet X.509 Public Key Infrastructure Certificate and Certificate Revocation List (CRL) Profile (RFC-5280)
• Santesson, S., Myers, M., Ankney, R., Malpani, A., Galperin, S., Adams, C., (2013) X.509 Internet Public Key Infrastructure a Certificate Status Protocol – OCSP (RFC-6960)
• Eastlake, D., (2011) Transport Layer Security (TLS) Extensions: Extension Definitions (RFC-6066)

Early Deliverables

1. Write a Java program which exhaustively solves the Travelling Salesman problem for small problems - say up to 10 cities.
2. Write a Java program which uses the nearest city hearistic and test its performance on large problems with 100-200 cities.
3. Write a report on these two algorithms which gives both asymptotic and actual runtimes for your implementations.
4. Write a Java implementation of simulated annealing which solves large Travellig Salesman problems.
5. Write a report describing the effects of modifying the parameters of your algorithm.

Final Deliverables

1. Write a report which gives a specification of a useful timetabling problem, and describes the software engineering approach that you will use to develop the final application.
2. Demonstrate a working timetabling system.
3. A final report on the project activities, fully acknowledging and citing sources.

Suggested Extensions

• Add graphical visualisations of your algorithms runtime activity and results.
• Consider the proble of rip-up and retry: taking an existing solution with new constraints and costs and attempting to produce a good modified solution that retains as much as possible of the old solution.

Reading

• Section on simulated annealing from the Numerical Recipes book published by Cambridge.
• Many online resources: start with Wikipedia

    AG (request => AF response)

Early Deliverables

1. Report: A survey of CTL and CTL model-checking algorithms.
2. Report: A survey of extended CTL.
3. Report: A survey of model-checking algorithms for context-free languages.
4. Proof of concept programs: A simple CTL model-checker.
5. Proof of concept programs: A simple analysis tool for context-free languages.

Final Deliverables

1. An analysis tool for finite-state systems and CTL extended with finite-state and context-free languages.
2. An experimental evaluation of the tool and comparison with CTL model-checkers.

Suggested Extensions

• Extensions allowing infinite-state systems to be analysed.

Reading

Early Deliverables

1. A report describing several TSP heuristics (e.g., Nearest Neighbour, Repeated Nearest Neighbour, Greedy).
2. Reports: Complexity, NP hardness and the big O notation
3. Implementations of Nearest Neighbour and Repeated Nearest Neighbour.

Final Deliverables

1. Description of several TSP heuristics (e.g., Nearest Neighbour, Repeated Nearest Neighbour, Greedy, Vertex Insertion).
2. Implementation of all heuristics described in the report.
3. Tables comparing quality and runtime of these heuristics on various testbeds.
4. GUI for the heuristics.
5. A report on TSP applications.
6. A short report on complexity of algorithms.

Suggested Extensions

• To study and implement Greedy+2-Opt. To include it into the tables.

Reading

• Start with Wikipedia on TSP.

Early Deliverables

1. A report describing several TSP heuristics (e.g., Nearest Neighbour, Repeated Nearest Neighbour, Greedy).
2. Reports: Complexity, NP hardness and the big O notation
3. Implementations of Nearest Neighbour and Repeated Nearest Neighbour.

Final Deliverables

1. Description of several TSP heuristics (e.g., Nearest Neighbour, Repeated Nearest Neighbour, Greedy, Vertex Insertion).
2. To study and implement Greedy+2-Opt.
3. Implementation of all heuristics described in the report.
4. Tables comparing quality and runtime of these heuristics on various testbeds.
5. GUI for the heuristics.
6. A report on TSP applications.

Reading

• Start with Wikipedia on TSP.

Early Deliverables

1. You will write a simple proof-of-concept program for computing Value at Risk for a portfolio consisting of 1 or 2 stocks using two different methods: model-building and historical simulation.
2. Different ways of computing Value at Risk will be back tested and the statistical significance of the results analyzed.
3. The report will describe the program (software engineering, algorithms etc.,).
4. The report will also discuss different ways of estimating the variance of returns and the covariance between returns of two different securities (using the standard formula, exponentially weighted moving average, or GARCH(1,1)).

Final Deliverables

1. The program should be extended by allowing derivatives (such as options) in the portfolio and using Monte Carlo simulation.
2. Allow an arbitrary number of stocks (using eigen- or Cholesky decomposition)
3. The final program will have a full object-oriented design, with a full implementation life cycle using modern software engineering principles.
4. Ideally, it will have a graphical user interface.
5. The final report will describe: the theory behind the algorithms.
6. The final report will describe: the implementation issues necessary to apply the theory.
7. The final report will describe: the software engineering process involved in generating your software.
8. The final report will describe: computational experiments with different data sets, methods, and parameters.

Suggested Extensions

• Computing conditional VaR (also known as expected shortfall)
• Explore carefully the choice of parameters for EWMA and GARCH(1,1) using a loss function such as square or log loss
• Empirical study of the two approaches to historical simulation for n-day VaR: reducing to 1-day VaR and direct estimation
• Complement back testing by stress testing
• Replacing the Gaussian model for stock returns by robust models
• Computing monetary measures of risk different from VaR

Reading

• John C. Hull. Options, futures and other derivatives, 7th ed., Pearson, Upper Saddle River, NJ, 2009.
• Hans Follmer and Alexander Schied. Stochastic Finance: An Introduction in Discrete Time, 3rd ed., de Gruyter, Berlin, 2011.
• Yahoo Finance as source of data: http://finance.yahoo.com/
• http://www.gloriamundi.org

Early Deliverables

1. You will write a simple proof-of-concept program for computing Value at Risk for a portfolio consisting of 1 or 2 stocks using two different methods: model-building and historical simulation.
2. Different ways of computing Value at Risk will be back tested and the statistical significance of the results analyzed.
3. The report will describe the program (software engineering, algorithms, etc.); discuss different ways of estimating the variance of returns and the covariance between returns of two different securities (using the standard formula, exponentially weighted moving average, or GARCH(1,1)).

Final Deliverables

1. The program should be extended by allowing derivatives (such as options) in the portfolio and using Monte Carlo simulation. Option pricing can be performed using the Black-Scholes formula, or binomial trees (for American options), or a different variety of Monte Carlo simulation (for Asian options).
2. Allow an arbitrary number of stocks in the model-building approach (using eigen- or Cholesky decomposition).
3. Allow an arbitrary number of stocks using histotical simulation.
4. Ideally, it will have a graphical user interface.
5. The final report will describe the theory behind the algorithms; the implementation issues necessary to apply the theory; the software engineering process involved in generating your software; computational experiments with different data sets, methods, and parameters.

Suggested Extensions

• Computing conditional VaR (also known as expected shortfall)
• Explore carefully the choice of parameters for EWMA and GARCH(1,1) using a loss function such as square or log loss
• Empirical study of the two approaches to historical simulation for n-day VaR: reducing to 1-day VaR and direct estimation
• Complement back testing by stress testing.
• Replacing the Gaussian model for stock returns by robust models.
• Computing monetary measures of risk different from VaR.
• Using proper scoring rules for quantiles for valuating computed VaRs.
• Using methods of worst-case risk aggregation, such as the Rearrangement Algorithm, or worst-case online decision making, such as the Weak Aggregating Algorithm (currently active areas of research).

Reading

• John C. Hull. Options, futures and other derivatives, 10th ed., Pearson, Upper Saddle River, NJ, 2018.
• Hans Follmer and Alexander Schied. Stochastic Finance: An Introduction in Discrete Time 3rd ed., de Gruyter, Berlin, 2011.
• Yahoo Finance as source of data: http://finance.yahoo.com/
• http://gloria-mundi.com

Early Deliverables

1. You will write a simple proof-of-concept program for computing Value at Risk for a portfolio consisting of 1 or 2 stocks using two different methods: model-building and historical simulation.
2. Different ways of computing Value at Risk will be back tested and the statistical significance of the results analyzed.
3. The report will describe the program (software engineering, algorithms, etc.); discuss different ways of estimating the variance of returns and the covariance between returns of two different securities (using the standard formula, exponentially weighted moving average, or GARCH(1,1)).

Final Deliverables

1. The program should be extended by allowing derivatives (such as options) in the portfolio and using Monte Carlo simulation. Option pricing can be performed using the Black-Scholes formula, or binomial trees (for American options), or a different variety of Monte Carlo simulation (for Asian options).
2. Allow an arbitrary number of stocks in the model-building approach (using eigen- or Cholesky decomposition).
3. Allow an arbitrary number of stocks using histotical simulation.
4. Ideally, it will have a graphical user interface.
5. The final report will describe the theory behind the algorithms; the implementation issues necessary to apply the theory; the software engineering process involved in generating your software; computational experiments with different data sets, methods, and parameters.

Suggested Extensions

• Computing conditional VaR (also known as expected shortfall)
• Explore carefully the choice of parameters for EWMA and GARCH(1,1) using a loss function such as square or log loss
• Empirical study of the two approaches to historical simulation for n-day VaR: reducing to 1-day VaR and direct estimation
• Complement back testing by stress testing.
• Replacing the Gaussian model for stock returns by robust models.
• Computing monetary measures of risk different from VaR.
• Using proper scoring rules for quantiles for valuating computed VaRs.
• Using methods of worst-case risk aggregation, such as the Rearrangement Algorithm, or worst-case online decision making, such as the Weak Aggregating Algorithm (currently active areas of research).

Reading

• John C. Hull. Options, futures and other derivatives, 10th ed., Pearson, Upper Saddle River, NJ, 2018.
• Hans Follmer and Alexander Schied. Stochastic Finance: An Introduction in Discrete Time 3rd ed., de Gruyter, Berlin, 2011.
• Yahoo Finance as source of data: http://finance.yahoo.com/
• http://gloria-mundi.com