A Deep Reinforcement Learning Framework for Contention-Based Spectrum Sharing

The increasing number of wireless devices operating in unlicensed spectrum motivates the development intelligent adaptive approaches to access. We consider decentralized contention-based medium access for base stations (BSs) on shared spectrum, where each BS autonomously decides whether or not transmit a given resource. contention decision attempts maximize its own downlink throughput, but rather network-wide objective. formulate this problem as partially observable Markov process with novel reward structure that provides long term proportional fairness terms throughput. then introduce two-stage time slot uses information from sensing and reception quality make decision. Finally, we incorporate these features into distributed reinforcement learning framework Our formulation inference, online adaptability also caters partial observability environment through recurrent Q-learning. Empirically, find maximization metric be competitive genie-aided energy detection threshold, while being robust channel fading small windows.