Martha White

The objective of our research program is to develop prediction and control systems that learn and adapt during deployment. For example, consider an agent whose goal is to improve efficiency of a water treatment plant. One of the primary costs is in the operation of mixers. The agent is constantly making decisions about mixing speeds, and receiving feedback about energy use and clean water production rates. We want this agent to be sample efficient, meaning that, with minimal interaction, it can learn about how to modulate mixing speeds appropriately. We additionally want the agent to continually learn (improve) during operation.

Towards this goal, we work in reinforcement learning. Specifically, my team focuses on: representation learning (sparse representations, recurrent architectures), planning with models based on options and value functions, off-policy learning to learn those model components, and better online control algorithms based on sound approximate policy iteration and action-value methods.