DS-PPO: A Reinforcement Learning Method for Wargaming

Wargame environments present unique challenges for reinforcement learning algorithms due to their large-scale state-action spaces, environmental asymmetry, and high random-ness. To address these challenges, this paper proposes DS-PPO, an improved reinforcement learning algorithm. DS-PPO incorpo-rates dual state input and action masking mechanisms to enhance the agent's perception of the environment and prevent invalid actions. Additionally, a redesigned loss function promotes policy ex-ploration and reduces instability. Experimental results demon-strate that DS-PPO outperforms other algorithms, achieving a win rate of 56.15%.