Philip Thomas

Philip Thomas's current research focuses on the mathematical foundations of computational theories of mind and their implications for artificial intelligence. He develops and analyzes candidate mathematical accounts of the relationship among physical systems, computation, experience, and valence. Given an explicit candidate account of experience or valence, he studies how AI systems can take agent well-being into consideration alongside task performance. This work includes qualia optimization and draws on machine learning, reinforcement learning, theoretical computer science, and philosophy of mind. It is centered in the Computational Minds and Agent Well-Being Laboratory (CMAW), which he directs.

Thomas also continues to work on reinforcement learning, particularly off-policy evaluation and safety. His earlier work includes high-confidence off-policy evaluation, Seldonian algorithms for safe and fair machine learning, and coagent methods for reinforcement learning. He directed the Autonomous Learning Laboratory (ALL) from 2017--2026 and remains a faculty affiliate.

Thomas co-founded the Reinforcement Learning Conference (RLC), a dedicated venue for reinforcement learning research. His work includes the Science article “Preventing Undesirable Behavior of Intelligent Machines,” and he has testified before the U.S. House of Representatives Task Force on Artificial Intelligence.

Thomas's research has received support from the National Science Foundation, the U.S. Army, DARPA, and companies including Insulet, Adobe Research, Dolby, Meta, Google, and Raytheon. Before joining the UMass Amherst faculty, he was a postdoctoral researcher with Emma Brunskill at Carnegie Mellon University.