Policy-Based Reinforcement Learning Approach in Imperfect Information Card Game

Games provide an excellent testing ground for machine learning and artificial intelligence, offering diverse environments with strategic challenges and complex decision-making scenarios. This study seeks to design a self-learning artificial intelligent agent capable of playing the trick-taking stage of the popular card game Thousand, known for its complex bidding system and dynamic gameplay. Due to the game's vast state space and strategic complexity, other artificial intelligence approaches, such as Monte Carlo Tree Search and Deep Counterfactual Regret Minimisation, are infeasible. To address these challenges, the enhanced version of the REINFORCE policy gradient algorithm is proposed. Introducing a score-related parameter beta designed to guide the learning process by prioritising valuable games, the proposed approach enhances policy updates and improves overall learning outcomes. Moreover, leveraging the off-policy experience replay, along with the importance weighting of behavioural policy, enhanced training stability and reduced model variance. The proposed algorithm was applied to the trick-taking stage of the popular game Thousand Schnapsen in a two-player setup. Four distinct neural network models were explored to evaluate the performance of the proposed approach. A custom test suite of selected deals and tournament evaluations was employed to assess effectiveness. Comparisons were made against two benchmark strategies: a random strategy agent and an alpha-beta pruning tree search with varying search depths. The proposed algorithm achieved win rates exceeding 65% against the random agent, nearly 60% against alpha-beta pruning at a search depth of 6, and 55% against alpha-beta pruning at the maximum possible depth.