Industrial <scp>IoT</scp> –Enabled Graph‐Structured Potential Game Learning With Communication‐Constrained Distributed Policy Gradients

ABSTRACT Industrial IoT–enabled cyber–physical systems increasingly require reliable and scalable decision‐making under graph‐coupled dependencies, partial observability, and constrained edge communication. Existing approaches often rely on centralized optimization or loosely coordinated multi‐agent learning, leading to limited robustness and inefficient communication. This paper proposes a communication‐constrained distributed learning framework based on graph‐structured potential games and policy gradient optimization. Each agent transforms streaming IoT observations into a compact risk‐aware state, enabling reliability‐oriented decision‐making without raw data exchange. A global potential function is constructed to align local utilities with system‐level reliability objectives while explicitly modeling topology‐induced externalities. To ensure stable learning under limited communication, agents exchange lightweight neighbor summaries and update policies via consensus‐regularized distributed gradients. Extensive experiments on benchmark distribution systems demonstrate that the proposed method achieves competitive reliability improvements with significantly reduced communication overhead, approaching centralized performance while maintaining scalability and robustness in dynamic environments.