Researchers at MIT and the University of Washington have devised a model for detecting the computational limitations of an agent, whether human or machine, that obstruct their ability to solve problems. Agents’ performance is monitored to calculate their “inference budget”, estimates of the time and effort likely to be re-invested in similar tasks, which then helps predict future behavior. In a recent paper, the investigators showed their model could deduce a person’s navigation goals from earlier routes and forecast the next moves of chess players. The technique either matched or exceeded the performance of other popular models.
The work has potential to advance AI-based systems understanding of human behavior and thus make them better co-workers. Predicting an individual’s goals from their actions could make an AI assistant more helpful, explains MIT student and study lead author, Athul Paul Jacob. “If we know that a human is about to make a mistake, having seen how they have behaved before, the AI agent could step in and offer a better way to do it,” he said.
The research will be unveiled at the International Conference on Learning Representations.
To create their model, the team took inspiration from previous studies of chess players. Their model can estimate an agent’s depth of planning from past actions and use that data to simulate the agent’s decision-making procedure.
Key to their process is running an algorithm for a limited period to solve the problem in question, such as playing a chess game for a certain number of steps. After comparing the algorithm’s decision-making process with that of the agent solving the same problem, the model is able to establish the agent’s inference budget.
This process is useful as the researchers can evaluate the whole range of decisions made by the decision-making algorithm without requiring additional work. The framework could be used anywhere a particular class of algorithms is applied.
The researchers applied their model in three modeling tasks: deducing navigation goals from earlier routes, guessing someone’s communicative objective from their verbal cues, and predicting the next moves in human-human chess matches. Their method either matched or exceeded a popular alternative in each trial. Furthermore, researchers found that their model of human behavior aligned well with measures of player aptitude in chess matches and task difficulty.
In the future, the team intends to use this method to model the planning process in other areas, such as reinforcement learning used commonly in robotics.
With the long-term objective of developing more effective AI collaborators, this study has been partly supported by the MIT Schwarzman College of Computing’s Artificial Intelligence for Augmentation and Productivity program and the National Science Foundation.
