Research On MEC Task Offloading And Resource Scheduling Based On Deep Reinforcement Learning

The maturity of 5G communication technology and the rapid development of the mobile Internet have made the scale of the Internet of Things(IoT)expand rapidly,becoming the key to the intelligence of many critical industrial and commercial systems and the opportunity for emerging technologies.While improving the quality of application services,emerging technologies have also brought many problems that need to be solved(such as insufficient computing power of mobile devices,difficulties in energy supply,and ultra-large-scale access to the IoT).Mobile edge computing(MEC)deploys resources to the edge network closer to users,providing mobile users with cloud-like functions that better meet users’ needs.At the same time,wireless power transmission(WPT)technology based on RF signals is considered a more efficient way than manually battery replacement or utilizing wired charging.Therefore,the combination of MEC and WPC technology solves the limitations of IoT devices in battery charging and computing power and provides a better environment for the development of IoT.So the research on WPT-MEC network architecture has important practical significance.This paper studies the problem of task offloading and resource scheduling in the WPT-MEC network.Through reasonable offloading decisions and resource allocation,the system can maintain the delay and energy consumption as low as possible and improve the quality of service.The main work of this paper is as follows:1.In a multi-user WPT-MEC network,a model is built for binary computing offloading and time resource allocation problems,maximizing transmission energy utilization while maximizing the mobile devices’overall computation rate.This paper proposes an online offloading and scheduling algorithm based on federated and deep reinforcement learning.Through reasonable offloading decisions and time allocation,the calculation rate of the mobile devices can be close to the optimal solution.In the case of non-iid data generated by mobile devices,an adaptive method is proposed to adjust the learning rate of neural networks to ensure the convergence and stability of the algorithm.Simulation results show that the algorithm is remarkably better than other baseline algorithms in terms of convergence speed,performance,and execution time and is close to the optimal solution with slight inaccuracy.2.On this basis,non-Orthogonal Multiple Access(NOMA)was introduced into the model.NOMA allocates the same communication resource block to multiple devices,which improves the spectrum efficiency,and effectively solves the problem of ultra-large-scale access to the IoT.Therefore,this paper proposes a WPT-MEC network architecture based on NOMA technology,which jointly optimizes the completion rate and consumption of computing tasks under system energy and delay limitation.Based on this problem,this paper first decoupled it into multiple subproblems,and adopted an algorithm of multi-round delay acceptance matching to solve the problem of sub-channel and transmission power allocation jointly.Then,a hybrid action deep reinforcement learning method based on TD3 and DQN is proposed,which jointly solves the problems of offloading decisions and MEC server computing resource allocation.Finally,simulation results show that the algorithm is better than the general baseline algorithm under the performance analysis of various indicators and is close to the optimal global solution.