Multi-Agent Explainable Federated Learning for Adaptive Resource Optimization in Intelligent Edge Computing Environments

Abstract

The rapid proliferation of Internet of Things (IoT) devices has shifted computation toward the network edge, where resource-constrained nodes must make fast, adaptive scheduling decisions under strict latency, energy, and privacy constraints. Existing edge resource-management approaches typically rely on either centralized reinforcement learning (RL), which requires raw data collection and therefore violates data privacy, or static heuristics that cannot adapt without manual tuning, and virtually none expose the reasoning behind their allocation decisions. This paper proposes MAEFL, a Multi-Agent Explainable Federated Learning framework that couples a federated (FedAvg-trained) demand-prediction surrogate with per-node multi-agent Q-learning policies and a permutation-importance explainability layer, enabling privacy-preserving, adaptive, and interpretable resource orchestration across heterogeneous edge nodes. We validate MAEFL on real telemetry from the Intel Berkeley Research Lab Wireless Sensor Network dataset — 12 physically distinct motes (edge nodes) and over 599,000 real, cleaned sensor readings — using each mote's genuine non-IID data distribution. On a real, ground-truth federated regression task (predicting real temperature from real humidity, light, voltage, and time-of-day), the federated model achieves a held-out R2 of 0.406 (MAE = 1.95°C) while exchanging 99.76% fewer bytes than centralizing raw readings. Driven by real trace-derived arrival and heterogeneity features, MAEFL reduces task-completion latency by 43.8% and improves resource-utilization efficiency by 37.2% relative to a non-adaptive Round-Robin baseline; a well-tuned Greedy-Least-Loaded heuristic and a privacy-violating centralized oracle both outperform MAEFL on raw latency/utilization by 7-15%, a genuine and discussed trade-off, with MAEFL's advantage lying in adaptivity without manual threshold tuning, privacy preservation, and explainability (fidelity ρ = 0.90). The contribution of this work is a jointly federated, multi-agent, and explainable resource-optimization framework, validated end-to-end on real IoT telemetry with an openly reproducible pipeline (including a Google Colab notebook) rather than purely synthetic simulation.