search

GAIGResearch / TabletopGames

MIT License
77 stars 55 forks source link

MCTS Max^N tree for multiple players #124

Closed hopshackle closed 3 years ago

hopshackle commented 3 years ago

Amend MCTS to support a vector of payoffs during back-propagation, with one value per player in the game. This can then be used to implement Max^N MCTS, in which each player makes a decision at their node using just their anticipated reward. See refs below.

Sturtevant, Nathan. 2008. ‘AN ANALYSIS OF UCT IN MULTI-PLAYER GAMES’. ICGA Journal, 14.

Sturtevant, Nathan, and Michael Bowling. 2006. ‘Robust Game Play against Unknown Opponents’. In Proceedings of the Fifth International Joint Conference on Autonomous Agents and Multiagent Systems - AAMAS ’06, 713. Hakodate, Japan: ACM Press. https://doi.org/10.1145/1160633.1160761.

hopshackle commented 3 years ago

All now implemented - see Pull Request #127