Progress in AI is best benchmarked by letting agents compete with each other. The Annual Computer Poker Competition (ACPC) was a powerful force for developing algorithms that can play Poker at a super human level.
During (2006 - 2018) ACPC greatly pushed algorithms for imperfect information games:
- CFR (2007)
- Solving limit Texas Hold’em Poker (2015)
- Finally, sound search methods (2017): we have computers bots that bested humans in no-limit Texas Hold’em Poker: DeepStack and Libratus.
ACPC significantly boosted research of imperfect information games, but ACPC is now dead. What games could now push the research forward?
In a spirit similar to ACPC, we’d like to organize a competition to test agents on large challenging games outside of Poker. We are specifically interested in large two-player zero-sum imperfect information games, which have some interesting and non-trivial structure of imperfect information.
- Move the science of imperfect information games forward.
- Use multiple games to drive generality.
- Community driven.
- Academy oriented.
- Have fun!
What we hope for
In some sense, large perfect-information games have been solved by the learning algorithm AlphaZero. As one of the primary goals of this competition we’d like to achieve a similar status for two-player zero-sum imperfect-information games: the same algorithm should be able to play any large game.
All decisions are agreed upon within organizing committee. The committee members are:
- Martin Schmid (DeepMind)
- Noam Brown (Facebook)
- Viliam Lisý (CTU)
- Michal Šustr (CTU)