Deep Neural Network Based Computational Resource Allocation for Mobile Edge Computing

The fifth generation of mobile technology, 5G, is facing a new challenge of explosive data traffic growth and massive device connection. 5G network new businesses such as driverless cars and smart grid require low delay, which are also energy-consuming applications. Mobile Edge Computing (MEC) is proposed as a new paradigm to provide computational resources for mobile users at the edge of mobile networks by deploying dense high-performance servers. Mobile devices (MDs) can migrate part of their tasks to the MEC server for parallel computation via wireless channel to obtain better user experience. Optimization algorithms have been reliable for solving such resource allocation problems. However, the iterative optimization algorithms are not suitable for the high real-time MEC system due to the complex operations and iterations. To tackle this challenge, we propose a deep neural network based algorithm. Firstly we use a classic optimization algorithm sequential quadratic programming (SQP) to get the optimization results. Then we train the DNN to approximate the behavior of SQP with the optimization results. The experiment results show that our proposed DNN based theme can be trained to well approximate SQP with high accuracy while speeding up the running time hundreds of times to meet the real-time requirement. Further, the comparison between the special DNN and general DNN show that we just need to train a general DNN with tolerable performance loss instead of training special DNNs towards different parameters like the number of MDs in the MEC system.