An Integrated Supervised–Reinforcement Learning Framework for Adaptive Decision-Making in Dynamic Environments

Abstract

Adaptive decision-making in dynamic environments is essential for intelligent systems such as autonomous robotics, smart healthcare, industrial automation, intelligent transportation, and cyber-physical systems. Traditional supervised learning models offer strong predictive performance when trained on labeled datasets but struggle to adapt to rapidly changing conditions and unseen scenarios. Reinforcement learning (RL) enables agents to learn optimal policies through continuous interaction with the environment; however, it often suffers from slow convergence, unstable training, and high exploration costs, limiting practical efficiency. To overcome these challenges, this study proposes an Integrated Supervised–Reinforcement Learning Framework (ISRLF) that combines both learning paradigms for robust and adaptive decision-making. Supervised learning is first used for feature extraction and knowledge initialization, providing a stable foundation for training. Reinforcement learning then refines decision policies through reward-based feedback from environmental interactions, enabling continuous adaptation. This hybrid approach improves learning efficiency, accelerates convergence, enhances adaptability, and reduces computational overhead. The framework incorporates deep neural networks for representation learning, policy optimization modules for decision refinement, environmental feedback loops for continuous learning, and adaptive reward mechanisms for intelligent control. Comparative evaluation against standalone supervised learning and RL models is performed using metrics such as accuracy, adaptability, convergence speed, response latency, and robustness. Results show that ISRLF achieves superior performance and stability in dynamic environments, making it suitable for next-generation real-time adaptive intelligent systems.