Reinforcement Learning-Based Intelligent Guidance Law for Air-to-Air Missiles

In the majority of research on guidance laws, missiles are commonly simplified as point masses. This approach may have adverse effects on the practical application of research outcomes. Additionally, conventional neural network-trained reinforcement learning (RL) guidance laws suffer from acceleration command oscillations (ACO) when the missile approaches the target, which can detrimentally affect the guidance performance. In order to tackle these issues, this study introduces a three-degree-of-freedom model that incorporates the missile body and controller in the training environment. Aiming to simulate the flight dynamics of missiles as closely to reality as possible. Furthermore, a gated recurrent unit layer is introduced in the neural network to allow the missile to leverage previous information and reduce chattering in output commands. Finally, the performance of the designed guidance law is validated through multiple repeated simulation experiments.