Dynamic Computation Offloading Based on Deep Reinforcement Learning

Mobile edge computing (MEC) provides computation capability at the edge of wireless network. To reduce the execution delay, computation-intensive multimedia tasks can be offloaded from user equipments (UEs) to the MEC server. How to allocate the computational and wireless resources is one of the key issues to guarantee the quality of services, and is very challenging when tasks are generated dynamically. In this paper, we address the above problem. To minimize the sum execution delay of multiple users, we jointly optimize the offloading decision and the allocation of both computational and wireless resources. We propose a deep policy gradient (DPG) algorithm based on the deep reinforcement learning. Simulation results show that our proposed DPG method can achieve lower latency than the baselines under different numbers of users, computation capacities and wireless bandwidths.