Abstract

Autonomous decision-making systems are transforming industries ranging from robotics and transportation to energy systems and manufacturing. At the core of this transformation are self-learning control architectures that enable systems to perceive environmental inputs, adapt internal models, and make decisions without direct human supervision. These architectures integrate machine learning, adaptive control theory, real-time feedback mechanisms, and often multi-agent coordination frameworks to achieve robust autonomy under uncertainty and dynamic conditions. This paper investigates state-of-the-art self-learning control architectures, exploring theoretical foundations, practical implementations, and performance outcomes across diverse application domains. Through an extensive literature review, empirical experiments, and comparative analysis, the research identifies key components of effective self-learning architectures, including reinforcement learning loops, neural adaptive controllers, and hybrid symbolic-statistical decision frameworks. The results indicate that self-learning architectures improve adaptability, resilience to unmodeled disturbances, and long-term performance in complex environments. However, challenges related to computational overhead, safety assurances, and interpretability persist. The paper concludes by presenting actionable design principles, quantified performance results, and a roadmap for future research in scalable, safe, and explainable self-learning control systems