Dynamic Multitarget Assignment Based on Deep Reinforcement Learning

Abstract

Dynamic multi-target assignment is a key technology that needs to be supported in order to improve the strike effectiveness during the coordinated attack of the missile swarm, and it is of great significance for improving the intelligence level of the new generation of strike weapon groups. Changes in ballistic trajectory during the penetration of multi-warhead missiles may cause the original target assignment scheme to no longer be optimal. Therefore, reassigning targets based on the real-time position of the warhead plays an important role in improving the effectiveness of the strike. In this paper, the dynamic multi-target assignment decision modeling method combining combat simulation and deep reinforcement learning was discussed, and an intelligent decision-making training framework for multitarget assignment was designed based on deep reinforcement learning. In conjunction with the typical combat cases, the warhead combat process was also divided into the penetration phase and the multi-target assignment phase, the model framework and reward function against the multi-target assignment of the missile were devised, and the SAC algorithm was employed to conduct application research on intelligent decision modeling for multi-target assignment. Preliminary test results suggest that the intelligent decisionmaking model based on deep reinforcement learning provides better combat effects than the traditional decision model based on knowledge engineering.INDEX TERMS Deep reinforcement learning, combat simulation, intelligent decision-making, multi-target assignment.

Military Decision Support with Actor and Critic Reinforcement Learning Agents

“…The WTA problem was solved by traditional optimization techniques and intelligent algorithms such as game theory, genetic algorithm, particle swarm algorithm, etc. 42 Wu, et al (2022) built the problem of multi-warhead penetration and striking multi-target using Python 42 . The defence strategy was trained using Soft Actor-Critic (SAC) and compared with the optimal rule model.…”