Offline Reinforcement Learning-Based Optimal Backoff Policy Selection for WSNs Using DQN: A Data-Driven Approach for Coexistence Management in the Unlicensed Spectrum

The unlicensed industrial, scientific, and medical (ISM) frequency bands are utilized by a plethora of heterogeneous wireless technologies that are densely collocated and share the same time and frequency constraints. With the ubiquitous connectivity paradigm of IoT and next-generation networks, an exponential proliferation of the ISM bands is expected and it will be a challenge to accommodate the growing number of heterogeneous ISM users whilst maintaining the quality of service of each wireless network. A novel data-driven coexistence mechanism for wireless sensor networks (WSNs) facing WiFi interference is proposed that uses Deep Q-Network (DQN)-based offline reinforcement learning for optimal policy selection of backoff values of WSN nodes as per their interference levels. Simulations show an up to 11.4% increase in network throughput as a result of the predicted optimal node backoff value as per the proposed algorithm. Furthermore, an exhaustive study of the node interference levels and node backoff values is performed and the subsequent effect on the network throughput and residual energy is studied. The optimal backoff policy formulation in the context of various network goals and applications is also discussed.