Supervisor: Johannes Fürnkranz, opens an external URL in a new window

Motivation:

Chess endgames with up to 7 men on the board have been completely solved in the sense that for every possible position, the best move, i.e., the move that wins in the minimum number of moves, can be looked up. However, these databases are very hard to understand for human players. On the other hand, many of these endgames have been analyzed by human players and simple strategies for playing them have been formulated.

Objective:

The goal of this thesis is to investigate on the example of a very simple endgame such as king-rook vs. king (KRK), whether simple strategies can be detected. The key challenge is to find a way that allows to go from the best move that is always avaiable in the database, to simple strategies which may consist of suboptimal moves, which nevertheless can be formulated in a concise way. Finding a good representation for a strategy is one of the challenges of the thesis. Possible ways to dectect such strategies could be learning from randomized game traces or from annotated human game traces.

Prerequisites

Basic knowledge in machine learning and chess endgames.