| dc.description.abstract | Stefan Dorra's For Sale is both a turn-based and simultaneous action zero-sum game where the objective is to become as rich as possible. The first phase of the game is a sequence of turn-based English auctions that bids for properties selected at random. The game itself is complex having a mix of multiple players, hidden information, and stochastic elements. Although auctions themselves have been thoroughly studied in literature this particular setup remains an open problem.
In this thesis, we investigate the usage of the interpretable Coalesced Tsetlin Machine (CoTM) for solving these types of auction games providing both excellent play and an understanding of how to play.
To this end, we first develop a self-playing reinforcement learning algorithm that achieves near optimal play. Secondly, based on this algorithm we construct a dataset with examples of optimal play. Thirdly, using CoTM we investigate various ways of understanding why particular moves are made.
The CoTM is also shown to outperform popular methods such as decision trees, neural networks, and k-nearest neighbours. On average the CoTM accuracy is 84.55\% significantly outperforming the other competitors.
We believe the resulting interpretability establishes that CoTM can be used for the interpretation of games that have a more complex game state than Hex and Go. | |
