Safe Reinforcement-Learning-Based Guidance Strategy Against Active Defense Target in Three-Player Engagement

In the three-player engagement, the missile must confront an active defense target—comprising both the target and the defender—while simultaneously pursuing the target and evading the defender. Existing intelligent missile guidance strategies against an active defense target integrate the reward for pursuing the target with the reward for evading the defender to improve guidance performance. However, setting the appropriate weights for these rewards poses a significant challenge. As a result, these strategies struggle to effectively optimize mission performance (i.e., pursuing the target) while simultaneously ensuring safety (i.e., evading the defender). Therefore, we propose a safe reinforcement learning-based guidance strategy against an active defense target in three-player engagement, enabling the missile to pursue the target while evading the defender efficiently. First, we model the kinematic and relative kinematic equations in a three-player engagement scenario. To apply safe reinforcement learning in the missile guidance against an active defense target, the Constrained Markov Decision Process framework for the missile is constructed. Next, from a probabilistic inference perspective, we propose a safe reinforcement learning-based guidance strategy in an Expectation Maximization style to efficiently guide the missile: the optimal non-parametric variational distribution is calculated using convex optimization (E-step), and policy parameters are updated through supervised learning (M-step). Finally, a wide range of simulation experiment results show the feasibility and superiority of the proposed guidance strategy in both static target and dynamic target environments.