Deep Reinforcement Learning With Bidirectional Recurrent Neural Networks for Dynamic Spectrum Access

Abstract

This paper considers the problem of Dynamic Spectrum Access (DSA) with partial observations under the cognitive radio framework. It is the heart of the matter for improving spectrum utilization efficiency. In general, DSA can be formulated as a partially observable Markov decision process (POMDP). While, extensive computational complexity is prohibitive for existing traditional methods. And the lack of environmental information makes some data-driven methods perform poorly. Bidirectional recurrent neural network (BRNN) is an extended form of unidirectional LSTM. Processing input sequence from both directions gives BRNN strong ability to perceive and predict sequential information. By introducing BRNN to deep reinforcement learning (DRL), Deep Bidirectional Recurrent Q-Network (DBRQN) is formed to obtain sequential correlation of input and achieve the optimal strategy. Extensive experiments indicate the efficiency and robustness of the proposed method in the case of simulated data and real data.