To create AI systems that can effectively collaborate with humans, having a solid model of human behavior is crucial. However, humans often make suboptimal decisions due to computational limitations.
The irrationality in human behavior, stemming from these constraints, is challenging to model. Humans cannot spend an extensive amount of time deliberating on a single problem.
Researchers from MIT and the University of Washington have developed a method to model the behavior of agents, whether human or machine, by considering the unknown computational constraints that may hinder their problem-solving abilities.
Their model can deduce an agent’s computational constraints by observing a few traces of their past actions. This information, known as the agent’s “inference budget,” can be used to predict their future behavior.
In a recent paper, the researchers demonstrate how their approach can infer someone’s navigation goals based on previous routes and forecast players’ next moves in chess games. Their technique either matches or surpasses another popular method for modeling this type of decision-making.
Ultimately, this research could aid in teaching AI systems how humans behave, enabling them to better interact with human partners. Understanding human behavior and inferring their goals from that behavior could enhance the usefulness of AI assistants, according to lead author Athul Paul Jacob, a graduate student in electrical engineering and computer science (EECS).
Jacob collaborated with Abhishek Gupta, assistant professor at the University of Washington, and senior author Jacob Andreas, an associate professor in EECS and a member of the Computer Science and Artificial Intelligence Laboratory (CSAIL). The findings will be presented at the International Conference on Learning Representations.
Modeling Human Behavior
Researchers have been developing computational models of human behavior for many years. Previous approaches often introduce noise to account for suboptimal decision-making. Instead of always selecting the correct option, the model might have the agent choose the right choice 95 percent of the time.
However, these methods may fail to capture the variability in human suboptimal behavior.
At MIT, other researchers have explored more effective ways to plan and infer goals despite suboptimal decision-making.
To construct their model, Jacob and his team drew inspiration from studies on chess players. They observed that players spend less time deliberating before making simple moves and that stronger players invest more time in planning during challenging matches.
“Ultimately, we found that the depth of planning, or the time someone spends thinking about the problem, is a reliable indicator of human behavior,” Jacob explains.
They developed a framework that could determine an agent’s planning depth from past actions and use that information to model the agent’s decision-making process.
The first step involves running an algorithm for a specific period to solve the problem under study. For example, in a chess match, the researchers might allow the chess-playing algorithm to run for a set number of steps. They can then analyze the decisions made at each step.
The model compares these decisions with the behaviors of an agent solving the same problem, identifying where the agent ceased planning.
From this, the model can establish the agent’s inference budget, or how long they will plan for the problem. This inference budget can predict how the agent would respond to a similar problem.
An Interpretable Approach
This method is efficient as it allows researchers to access the algorithm’s full set of decisions without additional effort. The framework can be applied to any problem solvable with a specific class of algorithms.
“What struck me the most was the interpretability of this inference budget. It indicates that more challenging problems require more planning, and stronger players plan for longer. Initially, we didn’t expect our algorithm to naturally pick up on these behaviors,” Jacob notes.
The researchers tested their method in three different modeling tasks: inferring navigation goals from past routes, deducing someone’s communicative intent from verbal cues, and predicting subsequent moves in human-human chess matches.
In each experiment, their approach either matched or outperformed a popular alternative. Additionally, their model of human behavior correlated well with player skill levels (in chess matches) and task complexities.
Looking ahead, the researchers aim to apply this approach to model the planning process in other domains, like reinforcement learning (commonly used in robotics). Ultimately, they plan to continue building on this work to develop more effective AI collaborators.
This research received support from the MIT Schwarzman College of Computing Artificial Intelligence for Augmentation and Productivity program and the National Science Foundation.
