Multi-task Deep Reinforcement Learning for Cognitive Spectrum-agile Communications

This paper introduces a cognitive engine design to achieve spectrum-agile communications over a heterogeneous wideband spectrum. The proposed cognitive approach has the ability to learn and avoid interference signals and other harmful signals. The targeted spectrum in this work is much wider than the ones proposed in the literature, most likely covering several hundreds of MHz. The proposed approach is based on deep reinforcement learning (DRL), more specifically on a double deep Q-network (DDQN) made of a convolutional neural network (CNN). The wideband spectrum is divided into a number of sub-bands and each sub-band consists of a number of channels. The problem is modeled as a multi-task DRL, where each sub-band represents a single task. Transfer learning is used between tasks to speed up the learning process. It is shown, through simulations, that the proposed technique can efficiently learn an effective strategy to avoid harmful signals in a noncontiguous wideband spectrum. Furthermore, it outperforms other DRL-based approaches in the literature while operating in a much wider spectrum and maintaining low computational complexity.