Learning-Based Distributed Resource Allocation in Asynchronous Multicell Networks

Abstract

A resource allocation problem is tackled in asynchronous multicell downlink LTE-LAA networks in pursuit of proportional fairness maximization by assuming limited channel state information (CSI) exchange. Previous studies solve resource allocation problems by relaxing the problems into fractional frequency resource allocation problems. Specifically, the binary resource allocation indicators are relaxed to real values, and the per-resource block (RB) signal-to-interference-plus-noise ratio (SINR) is averaged over all the RBs. However, the performance of such an approach is far beyond the optimality in frequency-selective channels. We propose a learning-based resource allocation framework only with limited CSI in frequency-selective channels. Without any additional CSI, we build a fully connected neural network architecture, based on which a distributed reinforcement learning algorithm is proposed. The proposed algorithm is implemented by using the TensorFlow library (Version 1.3.0 GPU) and python (Version 2.7). Numerical results show that the proposed learning-based algorithm exhibits enhanced proportional fairness performance compared to existing algorithms even with the same CSI assumption.