06-03-2017 11:30  Graduate Seminar

Solving Computer Poker with Counterfactual Regret Minimization

Counterfactual regret minimization (CFR) is an iterative algorithm that was developed in 2007 to solve for equilibrium strategies in imperfect information games. The algorithm is based on regret matching players reach equilibrium play by tracking regrets for past plays and making future plays proportional to positive regrets. We have seen significant advances in recent years including the solving of 1 on 1 Limit Texas Holdem Poker in 2015 and a poker agent defeating a group of four of the best poker players in the world in 1 on 1 No Limit Texas Holdem in 2017. We will look at how the algorithm works and what developments, which include monte carlo methods and abstraction, have lead to the impressive recent results. We will also look at experiments that compare different sampling and abstraction methods within CFR in a very small toy game and in a Texas Holdem-like game.

Location: 1061
Speaker: Max Chiswick
Affiliation: Dept. of Electrical Engineering Technion Back