基于边缘计算的任务卸载优化：深度强化学习方法

In recent years, with the booming development of modern electronic information technologies and industries, such as the Internet of Things (IoTs), big data, robots and artificial intelligence, mobile edge computing thus has urgent application scenario needs. To deal with the optimization problem on computing offloading and task scheduling, traditional methods like convex optimization, heuristic algorithm and reinforcement learning, are short at explosive growth challenges from increasing number of devices and date, user needs and changing scenarios. Fortunately, deep reinforcement learning is a good solution to this problem, which relies on deep neural network, and has become quite suitable for large state space. This paper is based on an edge computing and reinforcement learning platform. With regard to the three-layer fully connected structure of traditional neural network, nevertheless, such a fully connected neural network has to face some problems. For instance, low training efficiency, insufficient data filtering and over-fitting. Therefore, this dissertation has adopted the neural network structure of feed-forward and auto-encoder, tested and compared the training performance of parameters with different hidden layers and Activation Function. As a result, the convergence time can be shortened by 18.4%. The results of this paper would help researchers and operators to choose neural networks for the mobile edge computing optimizations.