Generative adversarial imitation learning computing task offloading scheme for optimizing of generated sample utilization and system overhead

An increasing number of studies have shown that generative adversarial imitation learning (GAIL) is an effective solution for multi-agent computing task offloading. However, the random policy assumption and model-free learning approach of GAIL result in low utilization efficiency of generated sample. In the rapidly changing battlefield, an edge environment lacking infrastructure and computational resources are limited, and the cost and overhead of agent-environment interaction are significant. It is not practical to collect a large number of samples for learning directly through continuous trial and error. This paper proposes a method for multi-agent computing task offloading method (MAGAIL-MCT) based on GAIL, utilizing mobile ad hoc cloud to integrate the computational resources of agents such as tanks, drones, and infantry. The offloading decision is evaluated through a scoring model and a policy iteration model, reflecting the advantages and disadvantages of offloading tasks to different terminal nodes. The experiment evaluates the algorithm’s performance based on the comprehensive system overhead and the utilization rate of generated sample, and explains the algorithm’s differences by analyzing changes in task quantity, decision time, offloading energy consumption, simulation duration, and mobility. Experimental results show that compared with MGAIL, DDPG-GAIL, BGAIL, and BC-GAIL, MAGAIL-MCT can minimize system overhead while improving sample utilization, reducing offloading delay, reducing offloading energy consumption, and mitigating the impact of mobility on method implementation, thereby enhancing the feasibility of intelligent computing task offloading on modern battlefields.