An intention-driven task offloading strategy based on imitation learning in pervasive edge computing

Abstract

Consider an infrastructure-less wireless network environment (e.g., a land battlefield) in which devices are characterized by varying resource configurations, dynamic mobility, complexity of the generated sensing tasks, and deterministic delay constraints for the processing of these tasks. Solving the associated problem is infeasible on many thin-client mobile or IoT devices. Existing research has not yet addressed the above issues. In this paper, we first analyze the latency problem that arises when offloading tasks to other neighboring devices for processing and model the self-benefit-maximizing task allocation process as a stochastic game. Second, by probing the state information of the available arithmetic resources, we model the problem of minimum Steiner tree (MST)-based task migration as a sequential decision-making process and construct a distribution of activity trajectories formed by the allocation decisions and state changes. Then, based on an expert system demonstration, multiagent imitation learning based on MSTs (MILMST) is proposed. For every task, the MST is used as the decision basis for task offloading based on the agents’ local observations, and the allocation strategy is gradually improved by interacting with the surrounding agents in an online manner. Finally, the superiority of our algorithm is experimentally demonstrated.