Deep Reinforcement Learning Based Task Offloading And Resource Allocation For Mobile Edge Computing

Emerging mobile applications,such as Augmented Reality/Virtual Reality,Internet of Vehicles,Artificial Intelligence,High-speed Video Streaming,require ultra-low service latency which is difficult to meet for traditional cloud computing architecture.By deploying computing resources at the edge,edge computing can process tasks nearby the user,which can alleviate the computing resource shortages of user equipment effectively and avoid data transmission with the remote cloud to reduce service delay.Edge computing has been one of the supporting technologies for the future network.However,compared with the remote cloud,the computing resources of edge cloud are still limited.Which tasks should be chosen to offload and How many resources should be allocated are becoming a research hot-spot.Due to the high execution complexity,it's hard for traditional optimization algorithms to apply in real-time scenarios.To solve this problem,a Deep Q Network based task offloading and resource allocation algorithm was proposed.The algorithm decomposes decision into sub-decisions to tackle the high dimension of state-action space,and defines system potential difference as reward function to learn the strategy that can minimize the weight sum of delay and energy.Besides,the mechanism of auxiliary reward was proposed to avoid the saturation state of resource allocation.Simulation experiments show that the proposed algorithm has a significant improvement in convergence,delay and energy consumption reduction and offloading coverage rate.In real computing scenarios,computing and storage are always inseparable.By caching user tasks at the edge,there is no need to transfer data with the edge cloud when tasks are requested again,which can further reduce latency and energy consumption.To solve the problem of cache-assisted task offloading and resource allocation,an active caching policy based on request pattern prediction and task volume awareness was proposed.This policy can effectively adapt to the evolution of task popularity and adjust the caching priority of tasks according to the task volume.Then,the DQN basedcache-assisted offloading and resource allocation algorithm was proposed.Simulation experiments verify the effectiveness of the proposed algorithm in request pattern prediction,caching policy and the consumption reduction of delay and energy.