Adaptive Computation Offloading Policy for Multi-Access Edge Computing in Heterogeneous Wireless Networks

Abstract

In heterogeneous wireless networks, massive mobile terminals randomly generate a large number of computation tasks (payloads). How to better manage these mobile terminals located in wireless networks to achieve acceptable quality of service (QoS) such as latency minimization, energy consumption minimization is crucial. A multi-access edge computing (MEC) server can be leveraged to execute the offloaded payloads generated from mobile terminals owing to its powerful processing power and location proximity features. However, an MEC server cannot tackle all offloaded tasks from multiple mobile terminals, and its energy consumption needs further consideration. We introduce an edge server model combined with the unmanned aerial vehicle (UAV) and equipped with the macro base station (MBS-MEC) to process the arrival payloads, and all UAVs and MBS-MECs can harvest renewable energy by using energy harvesting equipment. Furthermore, we model the computation offloading as a deep reinforcement learning scheme without <i>priori</i> knowledge. Considering the infeasibility of deep-reinforcement learning-based centralized learning for the proposed edge computing framework, we propose a distributed computation offloading scheme based on deep reinforcement learning (DCODRL) to minimize the weighted average cost, including the latency cost and the energy cost. Each mobile terminal can be regarded as a learning agent for the DCODRL. To compensate for the lack of cooperation of the DCODRL, we propose a gated-recurrent-unit-assisted multi-agent computation offloading scheme based on deep reinforcement learning (MCODRL) to improve the offloading policy by obtaining global observation information and designing a common reward for all agents. Comprehensive numerical results reflect the convergence and effectiveness of the DCODRL and MCODRL, and the efficacy of the proposed algorithms is further verified through comparisons with two baseline algorithms.