Generative Adversarial Imitation Learning-based Continuous Learning Computational Guidance

This paper proposes a generative adversarial imitation learning-based continuous learning computational guidance (GAIL-CLCG) to improve missile guidance capability. Conventional analytical guidance algorithms are usually unable to take into account dynamic changes such as drag and lift, and can only rely on simplified constant velocity model, which degrade performance under real-world condition. And existing computational guidance algorithms based on reinforcement learning face difficulty in reward function design. Our approach exploits the ability of generative adversarial imitation learning (GAIL) combined with gated progressive neural network (GPNN) to effectively address these issues. GAIL-CLCG directly generates guidance command by mimicking the behavior of expert, eliminating the need for elaborate human design of reward function. A distinctive feature of our approach is the incorporation of a GPNN, which supports continuous adaptation to new scenarios by leveraging prior knowledge. Simulation results on a large amount of data show that GAIL-CLCG not only successfully learns expert policy, but also improves the efficiency of adapting to different scenarios by migrating prior knowledge.