Q-learning based air combat target assignment algorithm

Target assignment is an important yet difficult problem in air combat. Previous methods, e.g., neural network, genetic algorithm, particle swarm optimization and ant colony algorithm for target assignment have been proved to be either too slow or not stable as far as converging to the global optimum is concerned. In this paper, Q-learning is verified to be an appropriate reinforcement learning algorithm for air combat target assignment. Firstly, the air combat Agents are modeled in term of their attributes, structure and actions; secondly, the criteria of state-action pairs are defined and the Q-learning based algorithm for target assignment is provided; in addition, the compromise between the exploration and exploitation of the algorithm is also discussed. Case analysis shows that the presented algorithm avoids relying on prior knowledge and performs well in getting out of local optimum.