MIT and University of Washington researchers have developed a model to understand and predict human behavior, which could improve the effectiveness of AI systems in collaboration with humans. Recognizing the suboptimal nature of human decision-making often due to computational constraints, the researchers created a model that factors in these constraints observed from an agent’s previous actions. This allows for the agent’s “inference budget,” which predicts their future behavior, to be automatically inferred.
In a new paper, the researchers demonstrated how their method could infer navigation goals from previous routes or predict subsequent moves in chess matches. This modeling technique matched or outperformed alternative methods popular in predicting decision-making outcomes. This research aims to teach AI systems to better understand and predict human behavior to optimize their responses, which would enhance their usefulness as AI assistants.
The team created their model by observing the planning depths of chess players, noting the correlation between planning time and match complexity or player skill level. The model infers an agent’s planning depth from previous actions and uses that information to predict the agent’s decision-making process. This enables the determination of the agent’s inference budget, which can be used to predict their reaction to a similar problem subsequently.
This model can be applied to any problem solvable by a specific class of algorithms, making it efficiently capable of accessing the full set of decisions made by the problem-solving algorithm. The researchers tested the approach on various tasks, including inferring navigation goals, guessing communicative intent, and predicting chess moves, recognizing a correlation between their model of human behavior and measures of task difficulty or player skill.
In the future, the researchers plan to model the planning process in other fields, such as reinforcement learning, with the ultimate goal of creating more effective AI collaborators. Funding support for the research came partially from the MIT Schwarzman College of Computing Artificial Intelligence for Augmentation and Productivity program and the National Science Foundation.
