Computationally constrained control in complex causal tasks
Abstract
This theory project aims to create new algorithms that choose smart actions for achieving goals. This requires new ways of building mental models of the world, while accounting for the difficulty of thinking. We call this computationally constrained control. Biological brains are able to interact efficiently with a complex world, and generalize far better than artificial agents to new situations while using vastly less energy than computer algorithms. Motivated by the strong performance of real brains in control tasks, we will generalize current theories of optimal control to include not only state rewards and action costs, but also the costs of thinking and the variability of neural machinery. We will design algorithms in accordance with the idea that good controls require good mental models that share some causal structure with the world. Our models will therefore use novel variants of machine learning tools that combine flexibility with structure. By analyzing the best performing algorithmic structures that minimize both task costs and thinking costs, we will identify fundamental principles of computationally constrained control that will be useful in a wide variety of practical applications, from the understanding and repair of brain function to the control of autonomous vehicles.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jan 21, 2022
- Source ID
- FA95502110422XX0
Entities
People
- Zachary Pitkow
Organizations
- Air Force Office of Scientific Research
- Baylor College of Medicine
- United States Air Force