Hybrid Deep Reinforcement Learning for Adaptive Decision-Making in Intelligent Control Systems
Indra Kishor, Udit Mamodiya, Mohammed Almaayah, Amer Alqutaish, Rami Shehab, Mansour Ratib Mohammad Obeidat · Engineered Science · 2025
Rapid and reliable decisions of intelligent control systems are more and more anticipated in uncertain environments with high dimensions where rule of thumb strategies can hardly be relied on. The traditional reinforcement learning models are often difficult to deal with issues of stability-adaptability trade-offs, particularly faced with time-varying sensory signals and slow feedback. To solve these issues, this paper proposes a Hybrid Deep Reinforcement Learning (HDRL) architecture that combines both deep representation learning and adaptive reward shaping as well as policy optimization of entropy. The architecture enables real time feedback alignment of both local exploration and global control goals which generates decisions which are stable even in noisy, dynamic environments. The validation of HDRL is achieved through experimental validation over benchmark control systems and has demonstrated that HDRL can achieve a maximum efficiency improvement of 15% on decision making, 12% faster convergence, and greater margins of stability than state-of-the-art hybrid models. It has been shown that the adaptive reward process coupled with self-regularized entropy tuning can result in the control policies that are less prone to variances and less energy-consuming. The work does not just have empirical benefits, but presents a conceptual connection between adaptive cybernetic feedback and autonomous development of control policies, hence offers a pathway to the development of more readable and self-organizing control intelligence. To the extent, the proposed framework can not only offer a scaled technical solution to this, but also an early change of how hybrid learning structures can offer stability and flexibility to intelligent control systems.