Task and Resource Allocation in Mobile Edge Computing: An Improved Reinforcement Learning Approach

Abstract

In this paper, the problem of joint task, power, and spectrum allocation is studied for a mobile edge computing (MEC) based small cell network (SCN). In the considered network, each small base station (SBS) that is connected with an MEC server provides commutation service to users while each MEC server provides computation service to users, who will request various types of computational tasks including: collaboration computational task performed at both the user and the MEC sever, edge computational task processed at the MEC server, and local computational task implemented by user itself. As the data size of each computational task requested by each user varies, one must reconsider resource (transmit power and subcarrier) allocation and task partition to effectively service the users. We formulate this problem as an optimization problem aiming to minimize the transmission and computational delay. To solve this problem, a multiple stack Q-learning method is proposed. The proposed method can use multiple stacks to record the wireless environment and users' information so as to avoid repeatedly learning the same information and, thus improving the learning efficiency and speeding up the convergence. Simulation results show that the proposed resource allocation scheme provides 79% and 63% gains in terms of sum delay compared to cases that fully computed at user and fully computed at the MEC server, respectively.