BiLSTM Based Reinforcement Learning for Resource Allocation and User Association in LTE-U Networks

Abstract

LTE-unlicensed (LTE-U) technology is a promising innovation to extend the capacity of cellular networks. The primary challenge for LTE-U is the fair coexistence between LTE systems and the incumbent WiFi systems. In this paper, we aim to maximize the long-term average per-user LTE throughput with long-term fairness guarantee by jointly considering resource allocation and user association on the unlicensed spectrum within a prediction window. We first formulate the problem as an NP-hard combinatorial optimization problem, then reformulate it as a non-cooperative game by applying the penalty function method. To solve the game, a novel reinforcement learning approach based on Bi-directional LSTM neural network is proposed, which enables small base stations (SBSs) to predict a sequence of future actions over the next prediction window based on the historical network information. It is shown that the proposed approach can converge to a mixed-strategy Nash equilibrium of the studied game and ensure the long-term fair coexistence between different access technologies. Finally, the effectiveness of the proposed algorithm is demonstrated by numerical simulation.