A Neuroevolutionary Approach for Opponent Modeling and Exploitation in No-limit Texas Hold’em Poker

In recent years, the research of solving incomplete information games such as Texas Hold’em Poker game promotes the development of Artificial Intelligence(AI), which includes a variety of advanced theories and approaches such as game theory, reinforcement learning, neural network, tree search and so on. As a landmark area in AI, interest in Texas Hold’em Poker game has never dropped off, since the breakthroughs in this area can be potentially applied to many real-world problems varying from financial negotiations to military confrontations. In this paper, we proposed a new framework with population-based neuroevolution for the problem of opponent modeling and exploitation in Texas Hold’em Poker. The framework composes of two core connected components: opponent network and game network. The opponent network is used to capture the opponents’ significant features that can be exploited and the pre-trained game network can thus make decisions accordingly. The game network is an end-to-end recurrent neural network and is evolved with a genetic algorithm in the training process. The feasibility of this scheme is verified through preliminary experiments in specific game scenarios.