Winning the “Nobel Prize” of computing
The world-changing AI research of Andrew Barto (PhD CSE ’75).
In early 2025 Andrew Barto (PhD CSE ’75) won a Turing Award, often referred to as the “Nobel Prize of Computing.”
The award recognizes the lasting impact of work from Barto and Richard Sutton on reinforcement learning, a machine learning paradigm that enables agents to make decisions through repeated interactions with their environment–essentially through trial and error. Reinforcement learning now forms the basis for widely applied AI technologies like Chat GPT.
The genesis for the world-changing work was a book Barto came across as an undergraduate in U-M’s Shapiro Undergraduate Library. Called “Brains, Machines, and Mathematics” by Michael A. Arbib, it was his first introduction to neural networks—a topic that would help define his career.
Barto’s interest in AI was also inspired by the work of John Holland, a professor in U-M’s Department of Computer and Communication Sciences (now Computer Science and Engineering). Holland was developing optimization algorithms—which today are called “genetic algorithms”—based on the trial and error of evolution by natural selection.
In genetic algorithms, each iteration represents a generation of solutions competing for survival. Top performers become “parents” that combine to create “children” with random mutations adding variation. Through repeated cycles, the process evolves toward optimal solutions, paralleling natural selection, where environmental pressures determine which individuals survive and reproduce.
The department’s biological slant led Barto to graduate school, where he spent five years modeling using cellular automata—mathematical systems where grids of simple cells follow local rules to produce complex, emergent behaviors.
After graduation, Barto continued research as a postdoctoral fellow at the University of Massachusetts Amherst, where he spent the rest of his career. There, Barto partnered with Sutton to explore combining neural networks and a selectional trial and error process into a system that could learn more like a human.
The approach was a major break with the conventional knowledge at the time, which favored a different approach called “supervised learning.” Supervised learning trains algorithms using labeled datasets—a bit like a teacher providing an answer key.
“In the early days, what we were doing was not fashionable at all. People were arguing that intelligence required symbolic processing and the work in the field was devoted there,” Barto said. “A few outliers were focusing on mathematical approaches, and that’s what I was interested in.”
Despite the controversy, Barto and Sutton eventually put their ideas into a textbook after many years of collaboration and learning from others in the field. “Reinforcement Learning: An Introduction” has now been cited more than 75,000 times.
Barto says that the far-reaching applications of reinforcement learning have come as a surprise. In addition, the increase in computational power has helped algorithms achieve things he couldn’t have imagined.
Barto is particularly pleased that reinforcement learning has helped optimize medical procedures for chronic diseases. However, in the flurry of AI developments, he encourages restraint.
“Reinforcement learning systems can learn to do things you didn’t expect, and sometimes those can be negative,” Barto said. “People need to be aware of that and use caution.”
Safety is a major research focus of the Autonomous Learning Laboratory at UMass Amherst—Barto’s former laboratory, which is now co-directed by two of his past students. The team works to let users flag dangerous or harmful behaviors and ensure that algorithms follow the rules.
“I’m glad to be retired and that I now have students who have taken over my lab or gone on to industry positions. It’s gratifying to see them doing great things,” Barto said.
One former student, Satinder Singh Baveja, even carries the torch at U-M, where he leads a reinforcement learning research group as a professor of computer science and engineering.
Andrew Barto and Richard Sutton will also jointly receive the 2026 IEEE Frank Rosenblatt Award, an award named in honor of one of the founders of neural networks.