Researchers from MIT and the University of Washington have developed a model that predicts human behavior by considering computational constraints that limit an individual’s problem-solving ability. This model can be used to estimate a person’s ‘inference budget’, or time available for problem-solving, based on their past actions. It can then predict their future behavior.
Drawing from research on chess players, the team built a system that can infer a person’s depth of planning from their previous actions. The method uses an algorithm to solve the given problem for a specific amount of time, such as the number of steps a chess algorithm could make within a set time limit. It then compares the decision made by this algorithm to an individual completing the same task. The model identifies at which point the individual ceased planning, establishing the ‘inference budget’. This can then be used to predict how they might approach similar problems in the future.
Electrical engineering and computer science graduate student and lead author Athul Paul Jacob cited examples of how such a model could be used. An AI tool could foresee a human about to make a mistake based on past behavior and suggest an alternative approach. Alternatively, the AI could adjust its functionality to accommodate the human’s shortcomings. This understanding of human behavior and the ability to predict future actions could vastly improve the usability and effectiveness of AI assistants.
This research forms part of a larger project focused on creating more effective AI collaborators. In particular, the team is interested in applying this approach to reinforcement learning which is commonly used in the field of robotics. Other applications for the model include predicting navigational routes and guessing communication intent from verbal cues.
The approach was successful in three different modeling tasks. It performed as well, if not better, than an alternative popular method in every instance. The interpretation of the human behavior was also consistent with measures of skill level and task difficulty.
The research, set to be presented at the International Conference on Learning Representations, is in part, financially supported by the MIT Schwarzman College of Computing Artificial Intelligence for Augmentation and Productivity program and the National Science Foundation.
