Liu, K.-H.K.-H.LiuWANJIUN LIAO2021-05-052021-05-05202015503607https://www.scopus.com/inward/record.url?eid=2-s2.0-85089430925&partnerID=40&md5=a46a04a59d40d6e3c332aba78b7847e2https://scholars.lib.ntu.edu.tw/handle/123456789/559290Multi-access Edge Computing (MEC) is promising to handle computation-intensive and latency-sensitive applications for 5G and beyond. Users can benefit from task offloading via wireless channels to MEC servers deployed at the nearby network edge. However, the radio resource is scarce and the computing resource in MEC is limited as compared to the remote cloud. Upon making an offloading decision, it is also important to efficiently allocate radio resource and MEC computing resource to ensure better service for the upload tasks. In this paper, we target the long-term delay and energy consumption performance in a multi-user system, and design an online solution based on Deep Reinforcement Learning (DRL) to deal with time-varying user requests and wireless channel conditions. To obtain better convergence property, we propose a new Actor-Critic model, called Discrete And Continuous Actor-Critic (DAC), to jointly optimize the continuous actions (i.e., radio resource allocation and computing resource allocation) and the discrete action (i.e., offloading decisions), and train the model iteratively with a weighted loss function. Our simulation results show that DAC outperforms existing solutions based on DDPG, DQN, and others, in terms of convergence speed, delay, and energy performance. © 2020 IEEE.Actor-Critic (AC); DAC; Deep Reinforcement Learning (DRL); Multi-access Edge Computing (MEC)[SDGs]SDG7Deep learning; Edge computing; Energy utilization; Radio; Reinforcement learning; Resource allocation; Actor critic models; Computation intensives; Convergence properties; Energy performance; Radio resource allocation; Sensitive application; Weighted loss function; Wireless channel condition; 5G mobile communication systemsconference paper10.1109/ICC40277.2020.91493872-s2.0-85089430925