CAREER: Discovering Theoretical Entities

University Of Maryland Baltimore County, Baltimore MD

Investigators

Abstract

The goal of this project is to develop algorithms that will allow robots to discover theoretical entities -- that is, causally relevant features of the environment that cannot be sensed directly. The algorithms developed by the PI will make it possible for robots to expand their understanding and control of the world around them by discovering theoretical entities, such as force and mass, in much the same way as human scientists do. This project also aims to shed light on human cognition, particularly about how we acquire foundational concepts. Bayesian Model Merging will be used to learn a model of the robot's environment. Non-determinism in action outcomes indicates the existence of theoretical entities, and mutual information among posited theoretical entities suggests that a single entity explains observed non-determinism and validates its causal efficacy. This work will be evaluated in three domains: (1) a simulated robot with realistic physics, (2) process control in an aluminum smelting plant, and (3) visual object recognition. This project will make it possible for robots to see beyond the "veil of perception", to overcome limitations of their sensory systems (as humans have), and to better understand, predict, and control their environments. In applying these algorithms to the aluminum smelting process, it will lead to reductions in energy consumption and production of greenhouse gases. Planned educational activities include working with female and minority undergraduate students, teaching a freshman seminar on the scientific method, and development of a new graduate-level course in robotics.