Learning-Based Resource Allocation: Efficient Content Delivery Enabled by Convolutional Neural Network

Abstract

In practical content delivery, when the time-frequency resources are limited, it is a challenging task to satisfy terminals' data demand in a heavy-traffic and mutual-interfered scenario. In this paper, we investigate time-efficient and energy-efficient solutions for content delivery at the network edge. We formulate two resource allocation problems, aiming at minimizing the total transmission time/energy in content delivery. The problems are formulated as mixed-integer linear programming. We obtain the global optimal solution by the branch-and-bound algorithm which typically incurs long computational time. To enable a computationally-efficient solution for fast and high-quality decision making, we resort to learning-based approaches to tackle the difficult combinatorial-optimization part. We investigate two deep-learning approaches, i.e., fully-connected deep neural network (FC-DNN) and convolutional neural network (CNN), to solve the problems. The FC-DNN and CNN are trained to learn and predict the discrete decisions. We compare the performance among FC-DNN, CNN, and the optimal solution. The numerical results illustrate that the proposed learning-based resource allocation approaches can achieve significant timesaving gains in computation and have promising performance in optimality approximation.