Distributed Computation Offloading in Mobile Fog Computing: A Deep Neural Network Approach
Zhongjun Yang, Wenle Bai
Abstract
In this letter, the performance of offloading is studied, which aims at minimizing the energy consumption in offloading with constraints on the delay. To solve binary computational offloading decision problem, a learning offloading algorithm based on distributed DNN is proposed, which uses multiple parallel DNNs to generate offloading decisions. The DNN are further improved by using the back-propagation method with cross-entropy as the loss function and using the newly generated decisions as a public training set. Besides, an innovative hierarchical offloading model is presented, based on which derives closed-form expressions for delay and energy in offloading. Then, the Delay-Energy Weighted Sum (DEWS) metric is defined and introduced as system utility to construct a gradient optimization problem to study. Extensive simulation indicate that the algorithm modified with the DEWS metric is able to achieve significant reduction on energy consumption with comparably less the total delay and maintain a high offloading accuracy.