Deep Reinforcement Learning in the Imperfect-Information Game Coup
2023
Hochschulschrift
Zugriff:
In recent years, deep reinforcement learning has been successful in many domains including two-player zero-sum imperfect-information games. However, most games tested with algorithms in this domain are small in terms of either the size of their state representation, the number of actions per game, or the overall size of their game tree. In addition, several games are abstracted or simplified versions of other games like poker.In this thesis, we apply deep reinforcement learning to the board game Coup. Coup is of particular interest because it centers around imperfect information and deception, and players benefit from using their memory to learn from the sequence of previous turns, which can inform their current strategy. Most importantly, Coup has a very large game tree, which will challenge existing algorithms. Deep CFR and NFSP are some of the most successful algorithms in this domain, and were selected to learn to play Coup. Several modifications to Deep CFR were required to make it more compatible with this large game. Specifically, we propose new game traversal sampling methods, and a more iterative variant of the algorithm. NFSP was able to perform significantly better than Deep CFR, though neither algorithm was able to achieve human-level performance. Evaluation of the trained agents in this large domain provided additional challenges. The agents’ performance was measured via several methods, including an approximation of exploitability, which estimates closeness to a Nash equilibrium. Overall, these experiments put a spotlight on certain strengths and weaknesses of existing algorithms in large games.
Titel: |
Deep Reinforcement Learning in the Imperfect-Information Game Coup
|
---|---|
Autor/in / Beteiligte Person: | Starcheus, Brandon |
Link: | |
Veröffentlichung: | 2023 |
Medientyp: | Hochschulschrift |
Schlagwort: |
|
Sonstiges: |
|