State‐Adaptive Personalized Actor–Critic for Human‐Centric Decision Path Optimization in Industrial <scp>IoT</scp> Systems

ABSTRACT Human‐centric decision optimization is increasingly important in Industrial Internet of Things (IIoT) systems and Industry 5.0, where heterogeneous devices, dynamic environments, and individualized human objectives must be jointly addressed. Conventional reinforcement learning (RL) methods often struggle with non‐stationary industrial conditions and diverse operator preferences. We propose a state‐adaptive personalized actor–critic framework for dynamic decision path optimization in IIoT environments. Unlike existing static‐policy or non‐personalized RL methods, the proposed framework jointly integrates user‐specific embeddings and recurrent state‐adaptive modulation within a unified actor–critic architecture. Policy learning, value estimation, and adaptive modulation are optimized through a unified objective with structured reward shaping. Experiments on benchmark path optimization and navigation datasets show up to 15% improvement in task success rate and 12% reduction in path length over strong RL baselines. The main contributions are: personalized actor–critic learning, recurrent state adaptation, and unified reward‐guided optimization. The framework provides a practical basis for robust human‐centric decision support in dynamic IIoT systems.