Construction of a Parallel Training Environment Model for Multi-Agent Deep Reinforcement Learning in Far-Sea Aerial Confrontation

• Published in: Hangkong Bingqi Vol. 32; no. 3; pp. 48 - 56
• Main Author: Zhang Yuan, Wang Jiangnan, Wang Wei, Li Xuan
• Format: Journal Article
• Language: Chinese
• Published: Editorial Office of Aero Weaponry 01.06.2025
• Subjects: far-sea region|aerial confrontation|multi-agent|deep reinforcement learning|jsbsim|training environment model
• ISSN: 1673-5048
• DOI: 10.12132/ISSN.1673-5048.2025.0020

The quality of the environment model determines whether the deep reinforcement learning system can efficiently and accurately learn and train to make good decisions. Aiming at the problems of idealized air combat environment construction and task scenarios in the context of far-sea and remote combat, this paper constructs a parallel training environment for multi-agent deep reinforcement learning in far-sea air combat. Among them, based on JSBSim and scalable radar and weapon system models, an agent model is built that takes into account both actual combat and simulation performance. This study selects 18-dimensional state space and 7-dimensional action space, and constructs a multi-reward system with the main line and 10 sub-objectives. This approach solves the problems of algorithm difficulty in convergence caused by poor guidance of sparse rewards and high dimensional space. The compliance of the environment, the effectiveness of classic deep reinforcement learning algorithms and compatibility with mainstr