Jon Kleinberg, Tisch University Professor in the departments of Computer Science and Information Science (in collaboration with colleagues from the University of Toronto and Microsoft Research), has undertaken research to better understand human decision making by studying how artificial intelligence (AI) plays chess. In a new paper, "Aligning Superhuman AI with Human Behavior: Chess as a Model System," the researchers study not how computers defeat humans when playing chess (AI's preeminence in that respect was confirmed decades ago), but how computers might be "trained to play like a human." As Kleinberg told Melanie Lefkowitz for her piece in the Chronicle:
“Chess sits alongside virtuosic musical instrument playing and mathematical achievement as something humans study their whole lives and get really good at. And yet in chess, computers are in every possible sense better than we are at this point,” Kleinberg said. “So chess becomes a place where we can try understanding human skill through the lens of super-intelligent AI.”
As Lefkowitz writes, "the researchers sought to develop AI that reduced the disparities between human and algorithmic behavior by training the computer on the traces of individual human steps, rather than having it teach itself to successfully complete an entire task. Chess—with hundreds of millions of recorded moves by online players at every skill level—offered an ideal opportunity to train AI models to do just that."
The research was supported in part by a Simons Investigator Award, a Vannevar Bush Faculty Fellowship, a Multidisciplinary University Research Initiative grant, a MacArthur Foundation grant, a Natural Sciences and Engineering Research Council of Canada grant, a Microsoft Research Award and a Canada Foundation for Innovation grant.
Read more in the Chronicle.
Abstract:
As artificial intelligence becomes increasingly intelligent—in some cases, achieving superhuman performance—there is growing potential for humans to learn from and collaborate with algorithms. However, the ways in which AI systems approach problems are often different from the ways people do, and thus may be uninterpretable and hard to learn from. A crucial step in bridging this gap between hu- man and artificial intelligence is modeling the granular actions that constitute human behavior, rather than simply matching aggregate human performance.
We pursue this goal in a model system with a long history in artificial intelligence: chess. The aggregate performance of a chess player unfolds as they make decisions over the course of a game. The hundreds of millions of games played online by players at every skill level form a rich source of data in which these decisions, and their exact context, are recorded in minute detail. Applying existing chess engines to this data, including an open-source implementation of AlphaZero, we find that they do not predict human moves well.
We develop and introduce Maia, a customized version of Alpha-Zero trained on human chess games, that predicts human moves at a much higher accuracy than existing engines, and can achieve maximum accuracy when predicting decisions made by players at a specific skill level in a tuneable way. For a dual task of predicting whether a human will make a large mistake on the next move, we develop a deep neural network that significantly outperforms com- petitive baselines. Taken together, our results suggest that there is substantial promise in designing artificial intelligence systems with human collaboration in mind by first accurately modeling granular human decision-making.