Researchers at MIT and the University of Washington have developed a method to effectively model human behavior, accounting for the computational constraints that limit our decision-making abilities. This model, known as the “inference budget,” enables predictions of an individual’s future actions based on their past behaviors. This is particularly useful in AI development, allowing machines to better understand human behavior and adapt to it.
The researchers showcased their method by predicting navigation goals based on previous routes and predicting subsequent chess moves. This technique matched or outperformed another popular method for modeling decision-making.
Understanding and predicting human behavior could make AI systems more efficient and effective at collaborating with humans. If an AI system can predict a mistake based on previous actions, it could intervene and provide a better solution. Additionally, the AI could adapt to the weaknesses of its human collaborators, further enhancing the AI-human partnership.
The team built their model by taking cues from how chess players strategize. They observed that players usually spend less time thinking before making simple moves while stronger players spend more time planning during challenging matches. By inferring the amount of planning based on previous actions, they modeled how that individual made decisions.
They put their model into practice using a problem-solving algorithm. The model compared the algorithm’s decisions with those of a human performing the same task and used that comparison to determine at what point the human stopped planning. This allowed the model to predict how that individual would behave when facing a similar problem.
This modeling method proved efficient and versatile, applicable to any problem that could be solved with a specific class of algorithms. The model’s ability to interpret levels of planning provided valuable insights into human behavior.
The team tested the model in three scenarios: inferring navigation goals from past routes, guessing intent from verbal cues, and predicting subsequent moves in chess matches. They found their method to either match or outperform a popular alternative in each experiment.
Looking ahead, researchers aim to apply this method to model planning processes in other areas, such as reinforcement learning techniques used in robotics. In the long run, this work is expected to contribute to the development of more effective AI-human collaboration. The MIT Schwarzman College of Computing Artificial Intelligence for Augmentation and Productivity program and the National Science Foundation funded this research.
