面向大机动目标的高效强化学习拦截制导律研究OA

In response to the challenges posed by intercepting highly maneuverable targets,including difficulties in obtaining target acceleration information,insufficient adaptability of traditional guidance laws to high-dynamic environments,low exploration efficiency of reinforcement learning algo-rithms in wide spatial-temporal contexts,and poor training stability,this paper proposes a terminal guidance law based on a double random distillation network and real proximal policy optimization algo-rithm.Firstly,a Markov decision process is designed for the three-dimensional terminal guidance sce-nario.Subsequently,deep reinforcement learning algorithms are employed to train high-speed vehicles.To enhance the exploration efficiency of these vehicles within wide spatial-temporal con-texts,a double random distillation network is introduced that incentivizes the exploration of unknown states through internal rewards.To accelerate training speed and ensure training stability,the trust region rollback objective function of real proximal policy optimization algorithm is used to replace the clipping objective function for effectively improving the convergence rate while enhancing training sta-bility.Simulation results demonstrate that the proposed terminal guidance law exhibits generalization and robustness when faced with highly maneuverable targets,and can achieve efficient interception without requiring target acceleration information.