Improving the Efficiency of Reinforcement Learning Using Model Predictive Control Solutions

Improving the Efficiency of Reinforcement Learning Using Model Predictive Control Solutions

Reinforcement learning (RL) algorithms are often employed in MRTs for the development of personalizing JITAIs (pJITAIs). These algorithms use participants’ accrued data to learn the optimal intervention decisions that are tailored to their state and context, and then adjust the randomization probabilities of delivering the interventions that could potentially maximize favorable health outcomes. Despite the usefulness of RL, one of the biggest challenges in healthcare applications is that the procedure is inefficient. An alternative way to tackle these issues is to use an approach from operations research, model predictive control (MPC). MPC is a powerful algorithm for learning the optimal controls (same as strategies) for mechanical systems.

Principal Investigator

Donglin Zeng, PhD
Professor, Biostatistics
School of Public Health
University of Michigan

Key Collaborators

Inbal Nahum-Shani, University of Michigan
Susan Murphy, Harvard University

Funding Source

Focus Area