Efficient Mastery of Gomoku Game Through Deep Monte Carlo Method

Deep reinforcement learning has made significant progress in the field of computer games. Traditional methods typically involve extensive tree expansion and simulation during the self-play phase to generate large amounts of match data for training models, which requires substantial computational resources. To reduce reliance on high computational resources, we abandoned the traditional self-play method and adopted a deep Monte Carlo method to generate match data in batches on GPUs, without relying on tree expansion. When applying this method to the game of Gomoku, the results showed that under limited computational resources, this method achieved a faster increase in Elo rating compared to traditional methods and exhibited faster convergence during training. Moreover, agents using this method ranked first among 503 agents on Peking University's BotZone platform.