Abstract:  Board games always have been a focus of deep reinforcement learning research due to their complex board configurations and rules， which require a lot of time to find optimal solutions. Current algorithms for chess games use action probability distribution-based methods for action selection during self-play， which leads to inefficient exploration and exploitation. They also require separate neural network computations for strategy and value， resulting in low sample usage and long training times. This paper proposes an efficient chess game algorithm that combines strategy-value networks， replacing the original action selection method with the Geng-Bellman maximum value method. It balances exploration and exploitation in action search using ε-greedy and simulated annealing algorithms. Experimental results show that compared to various classical chess game algorithms， the proposed algorithm achieves a win rate of over 90% against traditional algorithms. Moreover， using Gumbel-max method during training leads to significantly higher Elo ratings compared to traditional action selection methods with low Monte Carlo simulation counts. With training reaching 3000 Elo ratings， the proposed algorithm can save 50% of the time. 

Key words: board games, Monte Carlo tree search, Gumbel-max method, ε-greedy algorithm, simulated annealing algorithm ,