Games people – and machines – play: Unraveling strategic thinking to improve AI | MIT News

Gabriele Farina grew up in a small town in the mountainous wine region of northern Italy. Neither of his parents had a college degree, and even though they were both convinced they “didn’t understand math,” Farina said, they bought him the technical books he wanted and didn’t discourage him from going to a science-oriented school, rather than high school.

By the time he was 14, Farina had settled on an idea that would become the foundation of his career.

“I was very impressed with the idea that a machine could predict or make decisions better than humans,” he says. “The fact that man-made calculations and algorithms can create programs that, in some ways, surpass their creators, when we build on simple building blocks, has always amazed me.”

When he was 16, Farina wrote code to solve a board game he was playing with his 13-year-old sister.

“I used game by game to figure out the right moves and show my sister that she had lost long before we both realized it,” Farina said, adding that her sister was less than happy with her new system.

Now an assistant professor in the Department of Electrical Engineering and Computer Science (EECS) and principal investigator in the Laboratory of Information and Decision Systems (LIDS), Farina combines concepts from game theory with tools such as machine learning, optimization, and statistics to develop the theoretical and algorithmic foundations of decision making.

Enrolling at the Politecnico di Milano college, Farina studied automation and control engineering. However, over time, he realized that what piqued his interest was not just “using known techniques, but understanding and expanding their foundations,” he says. “I gradually switched to theory, while still deeply concerned with demonstrating the practical application of that theory.”

Farina’s advisor at Politecnico di Milano, Nicola Gatti, a professor and researcher in computer science and engineering, presented Farina with research questions on the theory of computer games and encouraged him to apply for a PhD. At the time, the first in her family to earn a college degree and living in Italy, where doctorates are treated differently, Farina says she didn’t even know what a PhD was.

However, one month after graduating, Farina began a doctorate in computer science at Carnegie Mellon University. There, he received awards for his research and writing, as well as a Facebook Fellowship in Economics and Computation.

While completing his doctorate, Farina worked for a year as a research scientist at Meta’s Fundamental AI Research Labs. One of his biggest projects was helping to develop Cicero, an AI that was able to beat human players in a game that involved making alliances, negotiating, and finding out when other players were wrong.

Farina says, “when we created Cicero, we designed it in such a way that it does not agree to make an alliance if it is not in its interests, and it also understands whether it is possible that the player is lying, because doing as they intended would be against their motives.”

2022 title in MIT Technology Review he said Cicero could represent progress towards AIs that can solve difficult problems that require compromise.

After his year at Meta, Farina joined the MIT faculty. In 2025, he was honored with the National Science Foundation CAREER Award. His work – based on game theory and its mathematical language that describes what happens when different groups have different goals, and then calls it an “equilibrium” where no one has reason to change their strategy – aims to simplify large, complex real-world situations where calculating such an equilibrium can take billions of years.

“I’m researching how we can use optimization and algorithms to get these very stable points,” he says. “Our work tries to shed new light on the basis of theoretical calculations, to better control and predict these complex dynamic systems, and uses these ideas to combine good solutions in large-scale multi-agent interactions.”

Farina is particularly interested in settings with “imperfect information,” meaning that some agents have information unknown to other participants. In those cases, information has value, and stakeholders must be strategic about using the information they have so as not to reveal it and reduce its value. An everyday example occurs in the game of poker, where players make mistakes to hide information about their cards.

According to Farina, “we now live in a world where machines are much better at making mistakes than people.”

A situation with “a large amount of incomplete information,” brought Farina back to his board game debut. Sttego is a military strategy game that has inspired multi-million dollar research efforts to produce systems capable of defeating human players. It requires accounting for complex risks and misunderstandings, or obfuscations, which may be the end of a classic game where great efforts have failed to produce superhuman performance, Farina said.

With new algorithms and training costing less than $10,000, instead of millions, Farina and his research team were able to beat the best player of all time – with 15 wins, four draws, and one loss. Farina says he is very happy to produce such results with such savings, and hopes that “these new methods will be integrated into the pipelines of the future,” he said.

“We’ve seen steady progress in building algorithms that can reason with each other and make sound decisions despite large action gaps or incomplete information. I’m excited to see these algorithms being integrated into the broader AI revolution happening around us.”