Model Aware Reinforcement Learning
Abstract
The overall goal in this proposal was to develop a novel model-aware RL (MARL) framework for nonlinear systems in continuous time and space specifically focused on mitigating large modeling errors and maintaining closed-loop stability during the learning phase. The original scope of work was to focus on the development of model-aware RL methods that utilize parametric models of the environment, are robust to modeling errors, and can adapt online to changing models and objectives. Data-driven adaptive estimation techniques were proposed to achieve online model estimation. Novel real-time model validation methods were proposed to gauge the quality of the estimated models. Development of fall-back policies was proposed to achieve robust learning in the presence of inaccurate models. In addition, the development model-aware RL methods that utilize non-parametric models such as Gaussian Processes (GPs) was proposed along with the use of the confidence bounds obtained for the GPs to guide model-aware virtual exploration. The development of model-aware RL techniques that utilize local parametric and non-parametric models was also proposed to synthesize locally optimal policies
Document Details
- Document Type
- Technical Report
- Publication Date
- May 01, 2023
- Accession Number
- AD1209329
Entities
People
- Rushikesh Kamalapurkar
Organizations
- Oklahoma State University–Stillwater