Targeted CFR

In recent years, Counterfactual Regret Minimization (CFR) has emerged as the standard technique for computing near-equilibrium solutions to large games of imperfect information. This paper describes a new sampling variant of Counterfactual Regret Minimization, called Targeted CFR. We compare with other sampling variants including Outcome Sampling and External Sampling, and present experimental results on poker. We find that Targeted CFR outperforms other sampling variants on certain types of large games.