To build AI systems that can effectively collaborate with humans, it helps to first have a good model of human behavior. However, humans tend to act suboptimally when making decisions.
These irrationalities, which are particularly difficult to model, often result in computational constraints. Humans cannot spend decades thinking of an ideal solution to a single problem.
Researchers at MIT and the University of Washington have developed a method to model the behavior of agents, whether human or machine, that accounts for unknown computational constraints that can hinder an agent's problem-solving ability.
The model can automatically infer the agent's computational constraints by looking at only a few traces of its previous actions. The results of an agent's so-called “inference budget” can be used to predict that agent's future behavior.
In a new paper, the researchers show how their method can be used to infer someone's navigation goal from their previous path and predict a player's subsequent moves in a chess match. Their technique matches or surpasses another popular method for modeling this type of decision making.
Ultimately, this research could help scientists teach AI systems how humans behave, which could make these systems more responsive to human collaborators. Being able to understand human behavior and infer goals from that behavior could make AI assistants much more useful, says Atul Paul Jacob, a graduate student in Electrical Engineering and Computer Science (EECS) and lead author of the paper. This technology.
“If it sees that a human is about to make a mistake by looking at how they have behaved before, an AI agent can step in and suggest a better way. Alternatively, the agent may adapt to the weaknesses of its human collaborators. “Being able to model human behavior is an important step toward building AI agents that can actually help humans,” he says.
Jacob co-wrote the paper with Abhishek Gupta, an assistant professor at the University of Washington, and senior author Jacob Andreas, an associate professor at EECS and a member of the Computer Science and Artificial Intelligence Laboratory (CSAIL). This research will be presented at the International Conference on Representations for Learning.
modeling behavior
Researchers have been building computer models of human behavior for decades. Many previous approaches try to account for suboptimal decisions by adding noise to the model. Instead of having an agent always choose the correct option, the model can ensure that the agent makes the correct choice 95% of the time.
However, these methods may not capture the fact that humans do not always do this. In the same way, act suboptimally.
Others at MIT have also studied more effective ways to plan and reason about goals in the face of suboptimal decisions.
To build the model, Jacob and his collaborators took inspiration from previous research on chess players. They found that when making simple moves, players took less time to think before acting, and that in difficult matches, stronger players tended to spend more time planning than weaker players.
“Ultimately, we learned that the depth of planning – how long someone thinks about a problem – is a good indicator of how humans behave,” says Jacob.
From their previous work, they built a framework that can infer an agent's planning depth and use that information to model the agent's decision-making process.
The first step in these methods is to run the algorithm for a set period of time to solve the problem under study. For example, if you are studying a chess match, you might want the chess game algorithm to run for a certain number of steps. Finally, researchers can see the decisions made by the algorithm at each step.
Their model compares these decisions to the actions of agents solving the same problem. Match the agent's decisions with those of the algorithm and identify the steps at which the agent stopped planning.
This allows the model to determine an agent's inference budget, or how long that agent will plan for this problem. You can use the inference budget to predict how that agent will react when solving a similar problem.
Interpretable Solutions
This method can be very efficient because it gives researchers access to the entire set of decisions made by the problem-solving algorithm without any additional work. This framework can also be applied to any problem that can be solved by a particular kind of algorithm.
“What was most surprising to me was that this inference budget was very interpretable. Difficult problems require more planning, and being a strong player means planning longer. “When we first started working on this, we had no idea that an algorithm would be able to capture this behavior naturally,” says Jacob.
The researchers tested their approach through three modeling tasks: inferring navigation goals from previous paths, inferring someone's communicative intentions from verbal signals, and predicting subsequent moves in a human-to-human chess match.
Their method performed as well or better than popular alternatives in each experiment. Moreover, the researchers found that their model of human behavior matched well with measures of player skill (in chess matches) and task difficulty.
In the future, the researchers want to use this approach to model planning processes in other areas, such as reinforcement learning (a trial-and-error method commonly used in robotics). In the long term, they plan to continue building on this work toward the larger goal of developing more effective AI collaborators.
This work was supported in part by the MIT Schwarzman College of Computing Artificial Intelligence for Augmentation and Productivity Program and the National Science Foundation.