An Energy-Efficient Multi-Agent Reinforcement Learning Approach for Spark Job Scheduling in Mobile Edge Computing
Battery life remains a major constraint on the continuous operation of mobile IoT devices represented by drones. Most existing computation offloading strategies rely on static heuristic rules, which fail to deliver stable performance in dynamically changing wireless mobile environments. Targeting this problem, this paper develops a multi-agent deep reinforcement learning approach for Spark job scheduling across local devices, edge servers and cloud resources. The scheduling task is modeled as a Markov decision process: six real-time system and network metrics, including CPU utilization, memory load, handover delay, signal strength, transmission latency and congestion level, form the state space, and the reward function is built directly on actual measured energy consumption data. Based on the multi-agent deep deterministic policy gradient (MADDPG) algorithm, the proposed method adopts a centralized training and decentralized execution paradigm to realize distributed collaborative decision-making. Validation on a public MEC dataset shows that the trained policy finally converges to a hybrid scheduling mode: 57% of tasks are processed locally, 39% are offloaded to edge nodes, and less than 4% are assigned to cloud resources for specific application scenarios. Comparative tests with PPO, FIFO, FAIR and HAS baselines confirm that multi-agent reinforcement learning can well capture the intrinsic scheduling patterns of complex mobile environments, providing an adaptive and energy-efficient scheduling solution for practical IoT deployments.