Researchers at MIT and the University of Washington have created a model that considers the computational constraints whilst predicting human behavior, which in turn could potentially make AI more efficient collaborators. These constraints can affect an individual or system’s problem-solving abilities. The model can automatically infer these constraints by observing only a few prior actions of the agent, also known as traces. This result is used to form an agent’s “inference budget”, which can then predict future behaviors.
Researchers have demonstrated their model’s capabilities by predicting navigation goals from prior routes and predicting players’ moves in chess matches.This new technique matches or outperforms another popular method for modeling this type of decision-making. Ultimately, understanding the behaviors of humans and inferring their goals from their actions can make AI assistants more helpful.
Furthermore, the team noted that players tended to think faster when making simple moves and stronger players spent more time planning in challenging matches. This observation led them to believe that the depth of planning is a significant indication of human behavior. They developed a framework that uses this insight to understand decision-making processes. The initial step in their method requires running a problem-solving algorithm for a fixed duration. The model then compares the decisions made at each step by the algorithm with those of an agent solving the same problem, aligns them and determines when the agent stopped planning.
This framework could potentially be applied to any problem that can be solved with a particular class of algorithm. The study found that their model of human behavior aligned well with player skills (in chess matches) and task difficulty. The team plans to use this approach to model planning processes in other areas, such as reinforcement learning, and to develop more efficient AI collaborators. The research was, in part, supported by the MIT Schwarzman College of Computing Artificial Intelligence for Augmentation and Productivity program and the National Science Foundation.
