Multi-Agent Edge AI for Coordinated Event Prediction over Multi-Layer Wireless Sensor Networks: A Smart Competitive Approach

Distributed event prediction over large-scale Wireless Sensor Networks (WSNs) requires balancing predictive accuracy against the cost of wireless communication. Centralized inference recovers strong accuracy but demands a continuously active uplink, while pure edge inference loses accuracy under partial observability. This paper proposes a communication-aware distributed inference framework that combines Echo State Network (ESN)-based local inference at edge agents with cluster-level majority voting, confidence fusion, and a confidence-gated upward communication rule across a three-level architecture of edge sensor agents, cluster-level agents, and a base station. A tunable confidence threshold θ, selected on a held-out validation set according to a named deployment objective, exposes three operating modes: High Sensitivity, Balanced, and High Efficiency. The empirical relationship between ensemble confidence Cavg and cluster decision accuracy—measured over ten confidence bins—confirms that high Cavg reliably predicts correct cluster decisions, providing direct empirical justification for using Cavg as the communication trigger. The framework is evaluated against a centralized SVM upper bound and a pure edge ESN lower bound across network sizes from 50 to 500 nodes on a synthetic nonlinear benchmark, with results averaged over five independent seeds. In Balanced mode, the multi-agent system achieves a global detection accuracy of 0.820–0.828 against 0.850 for the centralized SVM while reducing upward communication by 68–82% depending on network size. In High Efficiency mode, communication is reduced by 80–88% while maintaining precision above 0.88. A key scalability finding is that, at a fixed detection objective, the validation-selected threshold θ* increases with network size, yielding greater communication savings as the network grows without degrading detection accuracy.