Reinforcement learning for controlling continuous systems with uncertain dynamics and restricted states using robust neural dynamic programming

This dissertation presents a comprehensive investigation into the use of advanced neural network architectures and control methodologies for modelling and controlling complex robotic systems, with a primary focus on the Stewart Gough platform. Central to this work is developing and validating novel Differential Neural Networks (DNN) architectures designed to improve the fidelity of system identification in scenarios characterized by non-linear and time-varying dynamics. Through simulation and experimental validation, a new State-Input DNN (SIADNN) demonstrated superior identification accuracy over traditional DNN, particularly in capturing dynamic behaviours where system states non-trivially influence system responses. Beyond identification, the SIADNN architecture proves to be well-suited for applications in model predictive control and adaptive control frameworks, eliminating the need for extensive linearization and reducing computational burden. Its capability to model systems with time-varying parameters enables more robust and scalable solutions in control design. Additionally, the dissertation explores integrating the DNN model into reinforcement learning (RL) pipelines, leveraging the MATLAB Reinforcement Learning Toolbox to optimize control strategies. Experimental results confirm the efficacy of this hybrid approach in enhancing trajectory tracking and overall control performance of the Stewart platform. The synergy between neural identification and learning-based control highlighted in this work offers a robust framework for dealing with uncertain, high-dimensional robotic systems. The findings advocate for the broader adoption of hybrid neural architectures in intelligent control applications and pave the way for future research in real-time adaptive and data-driven control strategies.

https://orcid.org/0000-0002-7157-2052