Multi Agent Reinforcement Learning for Sleep Scheduling and Energy-Efficient Routing in Sustainable Wireless Sensor Networks

Abstract

Wireless sensor networks (WSNs) enable energyautonomous Internet of Things (IoT) services. However, intensive communications, idle listening, and non-coordinated sleep mechanisms drain batteries quickly and cause network unavailability. This study proposes GreenSense-MARL, a multiagent reinforcement learning framework that jointly optimizes adaptive sleep scheduling and energy-efficient data transmission in resource-constrained WSNs. Each sensor acts as an intelligent agent that perceives its residual energy, queue status, sensing variation, link quality, neighbor activity, and sink distance before choosing to take one of five actions: deep sleep, light sleep, sensing, transmit, or relay. Cooperative rewards promote collaboration among agents towards energy savings, packet delivery ratio, latency, sensing coverage, and duty-cycle fairness. Simulation results with over 2.3 million data records collected from 54 Mica2Dot sensor nodes in the Intel Berkeley Research Lab dataset show that GreenSense-MARL consistently outperforms EAP-IFBA, RL-DQN, and NCRDAP algorithms under the same simulation settings. GreenSense-MARL attained a packet delivery ratio, normalized network lifetime, and dutycycle fairness of 95.0%, 94.2%, and 93.6%, respectively, while capping average total energy consumption and average end-toend delay at 8.4% and 10.7%, respectively. Results show that decentralized learning algorithms effectively promote reliable, energy-aware, and sustainable communication schemes in dynamic IoT-enabled WSN deployments.