Cognition behavior model for air combat based on reinforcement learning

A cognition model was proposed to support tactical decisions for simulated fighters to fight with each other in a virtual combat, and reinforcement learning (RL) technology was used to acquire knowledge. The combat situation was described by multi-attributes, which resulted in a high dimensional problem space in which the fighters learned to find action policies. The traditional approach that partitioned the problem space would impose demand on huge computation and storage resource. An approximation network is constructed based on Gaussian radial basis function to approximate the state value, which greatly reduced the resource demand and learning cycle time, and produced reasonable maneuver strategy. The model was verified by a one-to-one air combat simulation, and the produced trajectories are similar with those that human pilots flied in real combat.