Air-Combat Strategy Using Deep Q-Learning

Unmanned aircraft systems (UAS) are essential components in the future air-combat. Due to high dynamics and randomness of the aircrafts, traditional methods are difficult to solve the optimal control strategy. The characteristics of reinforcement learning (RL) match the difficulty of this problem. In this paper, we build an air-combat game environment and train the agent with deep Q-learning (DQN). Despite of increasing probability of loses slightly, our method performs much better than other algorithms in the simulations. Compared with the searching based methods, like Minimax and MCTS, the policy trained by DQN can take specific tactics with a long-term view of the game. Result shows that a large number of simulations and carefully designed features and reward are the essential points of DON.