Application of Deep Reinforcement Learning in Maneuver Planning of Beyond-Visual-Range Air Combat

Beyond-visual-range (BVR) engagement becomes more and popular in the modern air battlefield. The key difficulty for pilots fight is maneuver planning, which reflects tactical decision-making capacity of both sides determinates success or failure. In this paper, we propose an intelligent planning method BVR combat with using improved deep Q network (DQN). First, a basic environment builds, mainly includes flight motion model, relative model missile attack model. Then, create decision framework agent interaction environment. Basic perceptive variables are constructed agents to form continuous state space. Also, considering threat each side constraint airfield, reward function designed training. Later, introduce training algorithm perceptional situation layers value fitting replace policy DQN. Based on long short-term memory (LSTM) cell, layer can convert high-dimensional perception situation. does well mapping action. Finally, three scenarios testing. Simulation result shows avoid enemy gather own advantages target. It also proves models methods valid be realized.