A Monte Carlo Neural Fictitious Self-Play approach to approximate Nash Equilibrium in imperfect-information dynamic games

摘要

Solving the optimization problem to approach a Nash Equilibrium point plays an important role in imperfect information games, e.g., StarCraft and poker. Neural Fictitious Self-Play (NFSP) is an effective algorithm that learns approximate Nash Equilibrium of imperfect-information games from purely self-play without prior domain knowledge. However, it needs to train a neural network in an off-policy manner to approximate the action values. For games with large search spaces, the training may suffer from unnecessary exploration and sometimes fails to converge. In this paper, we propose a new Neural Fictitious Self-Play algorithm that combines Monte Carlo tree search with NFSP, called MC-NFSP, to improve the performance in real-time zero-sum imperfect-information games. With experiments and empirical analysis, we demonstrate that the proposed MC-NFSP algorithm can approximate Nash Equilibrium in games with large-scale search depth while the NFSP can not. Furthermore, we develop an Asynchronous Neural Fictitious Self-Play framework (ANFSP). It uses asynchronous and parallel architecture to collect game experience and improve both the training efficiency and policy quality. The experiments with th e games with hidden state information (Texas Hold'em), and the FPS (firstperson shooter) games demonstrate effectiveness of our algorithms.