A Fusion Algorithm Based on HPSO and DDPG for Target Assignment Problem
Di Song, Li Yang, Qilong Hang · 2024
The target assignment problem has been encountered in many fields, such as games, industrial production, and military. We study how a set of defenders can be assigned to effectively intercept a set of attacking parties. The problem requires high precision in optimization and timeliness and stable solution performance, and randomly changing scenarios require assignment policies with stable performance and some generalization capabilities. This paper addresses this crucial issue by proposing a fusion algorithm combining Hybrid Particle Swarm Optimization (HPSO) with Deep Deterministic Policy Gradient (DDPG). Our contributions are as follows. First, the HPSO algorithm combines the Particle Swarm Optimization algorithm with local search methods to enhance its performance. We use HPSO to initialize DDPG to ensure the precision of the final assignment strategy. Second, we utilize the high-quality solutions generated by HPSO to reverse-generate experiences. Leveraging DDPG’s continuous learning capability and its fast and stable execution capability, we further enhance optimization precision while meeting the requirements of high timeliness and stable solving performance. Third, we design training scenarios with random variations to improve the model’s generalization ability. Numerical results indicate that the proposed method can rapidly make decisions and improve resource assignment.