Continuous-Time Reinforcement Learning: New Design Algorithms with Theoretical Insights and Performance Guarantees

This dissertation discusses continuous-time reinforcement learning (CT-RL) for control of affine nonlinear systems. Continuous-time nonlinear optimal control problems hold great promise in real-world applications. After decades of development, reinforcement learning (RL) has achieved some of the greatest successes as a general nonlinear control design method. Yet as RL control has developed, CT-RL results have greatly lagged their discrete-time RL (DT-RL) counterparts, especially in regards to real-world applications. Current CT-RL algorithms generally fall into two classes: adaptive dynamic programming (ADP), and actor-critic deep RL (DRL). The first school of ADP methods features elegant theoretical results stemming from adaptive and optimal control. Yet, they have not been shown effectively synthesizing meaningful controllers. The second school of DRL has shown impressive learning solutions, yet theoretical guarantees are still to be developed. A substantive analysis uncovering the quantitative causes of the fundamental gap between CT and DT remains to be conducted. Thus, this work develops a first-of-its kind quantitative evaluation framework to diagnose the performance limitations of the leading CT-RL methods. This dissertation also introduces a suite of new CT-RL algorithms which offers both theoretical and synthesis guarantees. The proposed design approach relies on three important factors. First, for physical systems that feature physically-motivated dynamical partitions into distinct loops, the proposed decentralization method breaks the optimal control problem into smaller subproblems. Second, the work introduces a new excitation framework to improve persistence of excitation (PE) and numerical conditioning via classical input/output insights. Third, the method scales the learning problem via design-motivated invertible transformations of the system state variables in order to modulate the algorithm learning regression for further increases in numerical stability. This dissertation introduces a suite of (decentralized) excitable integral reinforcement learning (EIRL) algorithms implementing these paradigms. It rigorously proves convergence, optimality, and closed-loop stability guarantees of the proposed methods, which are demonstrated in comprehensive comparative studies with the leading methods in ADP on a significant application problem of controlling an unstable, nonminimum phase hypersonic vehicle (HSV). It also conducts comprehensive comparative studies with the leading DRL methods on three state-of-the-art (SOTA) environments, revealing new performance/design insights.