Robust Coordinated Reinforcement Learning for MAC Design in Sensor Networks

In this paper, we propose a medium access control (MAC) design method for wireless sensor networks based on decentralized coordinated reinforcement learning. Our solution maps the MAC resource allocation problem first to a factor graph, and then, based on the dependencies between sensors, transforms it into a coordination graph, on which the max-sum algorithm is employed to find the optimal transmission actions for sensors. We have theoretically analyzed the system and determined the convergence guarantees for decentralized coordinated learning in sensor networks. As part of this analysis, we derive a novel sufficient condition for the convergence of max-sum on graphs with cycles and employ it to render the learning process robust. In addition, we reduce the complexity of applying max-sum to our optimization problem by expressing coordination as a multiple knapsack problem (MKP). The complexity of the proposed solution can be, thus, bounded by the capacities of the MKP. Our simulations reveal the benefits coming from adaptivity and sensors' coordination, both inherent in the proposed learning-based MAC.