Cognitive Reinforcement Learning Architecture for Autonomous Decision Optimization in Human–Machine Collaborative Environments

Abstract

Human–machine teams demand enhanced autonomy to improve task performance while reducing operator workload and unsafe interactions, but the design of such systems is challenging. This article proposes a Cognitive Reinforcement Learning Architecture (CRLA), which enhances the state representation of model-free reinforcement learning by incorporating estimates of workload, trust, fatigue, expertise, machine confidence and interaction history, and applies value learning and cognitive arbitration with a safety shield to obtain collaboration optimization. We assess the viability of CRLA as a shared control policy in a benchmark simulation testbed distilled from India-specific smart manufacturing, medical decision support and logistics use cases, against rule-based shared autonomy, Q-learning and human-in-the-loop variants, using 2,160 total evaluation episodes. CRLA achieved the maximum Collaboration Optimization Index (56.16) across three work domains, with a decrease in overload exposure of 11.29 percentage points compared to standard Q-learning and comparable mean rewards (p = 0.605). The highest gains were realized in smart manufacturing, while the high-stakes medical domain maintained higher levels of conservative verification. This study shows that cognition-aware, constrained autonomy can be a practical pathway to responsible human–AI decision-making in complex, safety-critical domains.