An End-to-End Intelligent Decision-Making Method for Air-Sea Collaborative Assault Operations

To address the challenges associated with decision-making in air-sea collaborative assault operations, this paper presents an end-to-end intelligent decision-making method based on deep reinforcement learning. In response to the issue of sparse rewards, a reward design approach that combines result rewards with process rewards is proposed. Furthermore, a preprocessing strategy for battlefield situation information is introduced to tackle the neural networks' challenge in comprehending complex battlefield information. The application of a state-dependent exploration mechanism aims to enhance the agent's efficiency in exploring the environment. Additionally, to accelerate model training, a centralized training-distributed sampling architecture is proposed. In the experimental section, data on average reward and winning rate demonstrate the effectiveness of the model proposed in this paper, including the reward function, preprocessing of battlefield situation information, and the state-dependent exploration mechanism. Moreover, the training speed data highlights the efficiency of the training architecture.