A Mission-Oriented Autonomous Missile Evasion Maneuver Decision-Making Method for Unmanned Aerial Vehicle

The aerial game environment is complex. To enhance mission success rates, UAVs must comprehensively consider threats from various directions and distances, as well as autonomous evasion maneuver decision-making methods for multiple UAV platforms, rather than solely focusing on threats from specific directions and distances or decision-making methods for fixed UAV platforms. Accordingly, this study proposes an autonomous missile evasion maneuver decision-making method for UAVs, suitable for multi-scenario and multi-platform transferable mission requirements. A three-dimensional UAV-missile pursuit-evasion model is established, along with state-space, hierarchical maneuver action space and reward function models for autonomous missile evasion. The auto-regressive multi-hybrid proximal policy optimization (ARMH-PPO) algorithm is proposed for this model, integrating autoregressive network structures and utilizing long short-term memory (LSTM) networks to extract temporal features. Drawing on exploration curriculum learning principles, temporal fusion of process and event reward functions is implemented to jointly guide the agent’s learning process through human experience and strategy exploration. Additionally, a proportion integration differentiation (PID) method is introduced to control the UAV’s maneuver execution, reducing the coupling between maneuver control quantities and the simulation object. Simulation experiments and result analysis demonstrate that the proposed algorithm ranks first in both average reward value and average evasion success rate metrics, with the average evasion success rate approximately 8% higher than the second-ranked algorithm. In the three initial scenarios where the missile is positioned laterally, head-on, and tail-behind the UAV, the UAV’s missile evasion success rates are 95%, 70%, and 85%, respectively. Multi-platform simulation results demonstrate that the decision model constructed in this paper exhibits a certain degree of multi-platform transferability.