or
Looking to list your PhD opportunities? Log in here.
About the Project:
The way in which AI systems make decisions is often different from the approach a human takes. Thus, the operation of an AI system may be difficult to interpret and learn from, and poses the problem of user trust. Many state-of-the-art models (e.g. deep neural networks) operate as black boxes optimised for predictive accuracy, with no provision for interpretability of the results – a property desired in scenarios like medical diagnoses or security systems. Further, when a learning-based system achieves superhuman performance, it is desirable to understand the patterns learned by the system as they may improve human understanding of the problem itself.
Chess has historically been a platform and a benchmark in AI research. Recently, methods based on deep learning for position evaluation (AlphaZero) emerged, showing state-of-the-art performance. Compared to traditional chess engines (e.g. Stockfish), they evaluate a drastically smaller number of positions, relying on learning a better position evaluation function with a DNN, presumably capturing higher-level, more human-like “positional understanding”.
However, the discrepancy between human reasoning and chess engines is apparent: humans cannot remotely compare with chess engines in the number of positions they evaluate, and rely on higher-level understanding of the game and heuristics with which to conceptualise and guide analysis.
Evidence from image-based and other types of DNN models suggests that their performance can be attributed to building incrementally higher-level and more abstract representations of data. We hypothesise that chess can also be modelled in a similar way and that chess is a promising platform for general interpretable AI research. This project will concentrate on devising inherently interpretable DNN models for chess, and on their analysis and generalisation to other interpretable AI problems. To this end the project will:
(1) Investigate whether it is possible to train models to recognise and exploit low-level tactical patterns (e.g. “pin”, “fork”, “X-ray”) and use these tactical models as inputs to a chess engine, instead of raw game states.
(2) Investigate training DNN models that encapsulate the highest, strategic levels of abstraction in chess (e.g. “controlling the centre”, “locking the position”, “war of attrition”).
(3) Develop a complete DNN evaluation engine designed explicitly to reason at multiple levels of abstraction, with potential to learn previously unknown mid-level abstractions in chess.
(4) Develop practical tools for human-interpretable analysis of chess and for training of human players. Generalise results to other decision-making systems.
Contact for more information on the project: Dr Kirill Sidorov; sidorovk@cardiff.ac.uk
Academic criteria: A 2:1 Honours undergraduate degree or a master's degree, in computing or a related subject. Applicants with appropriate professional experience are also considered. Degree-level mathematics (or equivalent) is required for research in some project areas.
Applicants for whom English is not their first language must demonstrate proficiency by obtaining an IELTS score of at least 6.5 overall, with a minimum of 6.0 in each skills component.
How to apply:
Please contact the supervisors of the project prior to submitting your application to discuss and develop an individual research proposal that builds on the information provided in this advert. Once you have developed the proposal with support from the supervisors, please submit your application following the instructions provided below
Please submit your application via Computer Science and Informatics - Study - Cardiff University
In order to be considered candidates must submit the following information:
If you have any questions on the application process, please contact COMSC-PGR@cardiff.ac.uk
The university will respond to you directly. You will have a FindAPhD account to view your sent enquiries and receive email alerts with new PhD opportunities and guidance to help you choose the right programme.
Log in to save time sending your enquiry and view previously sent enquiries
The information you submit to Cardiff University will only be used by them or their data partners to deal with your enquiry, according to their privacy notice. For more information on how we use and store your data, please read our privacy statement.
Research output data provided by the Research Excellence Framework (REF)
Click here to see the results for all UK universitiesBased on your current searches we recommend the following search filters.
Check out our other PhDs in Cardiff, United Kingdom
Start a New search with our database of over 4,000 PhDs
Based on your current search criteria we thought you might be interested in these.
Developing the next generation of pedestrian behaviour models for revival of high streets and sustainable transport [Self-Funded Students Only]
Cardiff University
Privacy risks and countermeasures for IoT devices [SELF-FUNDED STUDENTS ONLY]
Cardiff University
Efficient and scalable consensus algorithms for decentralised cryptocurrencies [SELF-FUNDED STUDENTS ONLY]
Cardiff University