Multi-Agent Systems on Sensor Networks: A Distributed Reinforcement Learning Approach

Implementing a multi-agent system (MAS) on a wireless sensor network comprising sensor-actuator nodes with processing capability enables these nodes to perform tasks in a coordinated manner to achieve some desired system-wide objective. In this paper, several distributed reinforcement learning (DRL) algorithms used in MAS are described. Next, we present our experience and results from the implementation of these DRL algorithms on actual Berkeley motes in terms of communication, computation and energy costs, and speed of convergence to optimal policies. We investigate whether globally optimal or merely locally optimal policies are achieved. Finally, we discuss the trade-offs that are necessary when employing DRL algorithms for coordinated decision-making tasks in resource-constrained wireless sensor networks.