CFR: Counterfactual Regret Minimization

A Python library implementing various counterfactual regret minimization algorithms for imperfect information games.

Disclaimer

This library is currently in a pre-alpha stage. All code is subject to change without prior notice.

Roadmap

The aim is to make a fully generic self-play algorithm that can be configured with the following parameters

• Update methods:
  • CFR
  • CFR+
  • Fictitious play
  • Hedge
• Graph methods:
  • Full width ("vanilla")
  • Chance sampling ("Monte Carlo")
• Optimizations:
  • Fixed strategy iteration
  • Pruning
• Games:
  • Rock-Paper-Scissors
  • Liar Die
  • (optional, no promises!) any game in Gambit EFG format through its PyAPI
  • (long-term goal) Stratego

Acknowledgments

This library was inspired by Todd W. Neller's and Marc Lanctot's lecture notes "An Introduction to Counterfactual Regret Minimization". Their lecture notes, slides and original Java implementations for the games of Rock-Paper-Scissors, Kuhn poker, Dudo and Liar Die have been added to this repository for archival purposes.

License

Copyright Rein Halbersma 2018.
Distributed under the Boost Software License, Version 1.0.
(See accompanying file LICENSE_1_0.txt or copy at http://www.boost.org/LICENSE_1_0.txt)