Deep Reinforcement Learning-based Behaviour Generation Algorithm for Air Combat Escape Intention

Although deep reinforcement learning applied to air combat has achieved good results, it still faces a series of challenges such as reward design, convergence of suboptimal solutions, and poor stability. In this regard, this paper proposes a behaviour generation algorithm based on Dueling-Noisy-Multi-step DQN for air combat under escape intent. By analysing the air combat confrontation process, we extract the escape intention features and establish the corresponding reward model; for the problem of poor stability and slow convergence of deep reinforcement learning algorithms in large-scale state-action space, we propose the Dueling-Noisy-Multi-step DQN algorithm, which improves the accuracy of the value function fitting and at the same time increases the efficiency of spatial exploration and network generalization. Comparison with other algorithms through simulation experiments, the results reflect the excellent performance of this paper's algorithm.