Researchers at MIT and the University of Washington have developed a new way to model human behaviour by accounting for unknown computational constraints that may impact problem-solving abilities. This new model enables an agent, human or machine, to infer another agent’s computational constraints from their previous actions. The resulting ‘inference budget’ can be used to predict an agent’s future behaviour. The researchers believe this could help in teaching AI systems how humans behave, making them more effective collaborators.
This method was outlined in a new paper and demonstrated its efficacy through inferring navigation goals from prior routes and predicting subsequent moves in chess matches. Importificantly, this method matched or even outperformed another popular method used for modelling this type of decision making.
“Being able to model human behavior is an important step toward building an AI agent that can actually help that human,” says Athul Paul Jacob, lead author of the paper. The researchers managed to build the model by studying chess players, discovering that the depth of planning or time taken to think about a problem was a good indicator of human behavior.
The first step in their model involves using an algorithm for a set amount of time to solve a problem. This algorithm’s decisions are then compared with the decisions of an agent solving the same problem. This comparison helps ascertain when the agent stopped planning, and then identifies the agent’s inference budget or how long the agent will continue to plan for this problem.
The model was tested in three different tasks – inferring navigation goals, guessing communicative intent from verbal cues, and predicting the next moves in chess games. The method either matched or outperformed another popular alternative in each case.
This research is supported by the MIT Schwarzman College of Computing Artificial Intelligence for Augmentation and Productivity program and the National Science Foundation. The team plans to extend the model to other domains, with an ultimate goal of developing more effective AI collaborators.
