Deep Counterfactual Regret Minimization in Continuous Action Space

Abstract

Counterfactual regret minimization based algorithms are used as the state-of-the-art solutions for various problems within imperfect-information games. Deep learning has seen a multitude of uses in recent years. Recently deep learning has been combined with counterfactual regret minimization to increase the generality of the counterfactual regret minimization algorithms. This thesis proposes a new way of increasing the generality of the counterfactual regret minimization algorithms even further by increasing the role of neural networks. In addition, to combat the variance caused by the use of neural networks, a new way of sampling is introduced to reduce the variance. These proposed modifications were compared against baseline algorithms. The proposed way of reducing variance improved the performance of counterfactual regret minimization, while the method for increasing generality was found to be lacking especially when scaling the baseline model. Possible reasons for this are discussed and future research ideas are offered.