To build artificial intelligence systems that can effectively cooperate with humans, it is worth starting with a good model of human behavior. But people tend to behave suboptimally when making decisions.
This irrationality, which is particularly arduous to model, often comes down to computational limitations. A person cannot spend decades thinking about the perfect solution to one problem.
Researchers at MIT and the University of Washington have developed a way to model the behavior of an agent, whether human or machine, that takes into account unknown computational constraints that can hamper an agent’s problem-solving ability.
Their model can automatically infer an agent’s computational limits by seeing just a few traces of its previous actions. The result, the so-called an agent’s “inference budget” can be used to predict the agent’s future behavior.
In the fresh paper, the researchers show how their method can be used to infer navigation goals from previous routes and predict players’ next moves in chess matches. Their technique equals or exceeds other popular methods for modeling this type of decision making.
Ultimately, this work could support researchers teach artificial intelligence systems how humans behave, which could enable these systems to better respond to their collaborators. The ability to understand human behavior and then infer its goals from that behavior could make an AI assistant much more useful, says Athul Paul Jacob, an electrical engineering and computer science (EECS) graduate student and lead author of the book article about this technique.
“If we know a human is about to make a mistake, after seeing how it has behaved before, an AI agent can step in and suggest a better way to do it. Or the agent could adapt to the weaknesses of its human associates. The ability to model human behavior is an essential step towards building an artificial intelligence agent that can actually support that person,” he says.
Jacob 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). The research results will be presented at the International Conference on Learning Representations.
Behavior modeling
Scientists have been building computational models of human behavior for decades. Many previous approaches attempt to account for suboptimal decision-making by adding noise to the model. Instead of having the agent always choose the correct option, the model can make the agent make the correct choice 95 percent of the time.
However, these methods may not take into account the fact that people do not always behave in the same suboptimal ways.
Others at MIT have also explored more effective ways to plan and reason about goals in the face of suboptimal decision-making.
To build their model, Jacob and his colleagues drew inspiration from previous studies of chess players. They noticed that players spent less time thinking before acting when making uncomplicated moves and that stronger players tended to spend more time planning than weaker ones in arduous matches.
“Ultimately, we found that the depth of planning, or how long someone thinks about a problem, is a really good indicator of people’s behavior,” Jacob says.
They built a framework that can infer an agent’s planning depth from previous actions and apply this information to model the agent’s decision-making process.
The first step in their method involves running the algorithm for a specified period of time to solve the problem under study. For example, if they are learning a chess match, they may let the chess algorithm run for a certain number of steps. Finally, researchers can see the decisions made by the algorithm at every step.
Their model compares these decisions with the behaviors of an agent solving the same problem. It will align the agent’s decisions with the algorithm’s decisions and identify the step at which the agent stopped planning.
Based on this, the model can determine the agent’s inference budget or how long the agent will plan to solve this problem. It can apply the inference budget to predict the agent’s response when solving a similar problem.
An interpretable solution
This method can be very capable because researchers can access the full set of decisions made by the problem-solving algorithm without performing additional work. This framework can also be applied to any problem that can be solved by a specific class of algorithms.
“For me, what was most striking was the fact that this application budget is very straightforward to interpret. This means that harder problems require more planning, and being a sturdy player means planning longer. When we started doing this, we didn’t think our algorithm would be able to naturally capture these behaviors,” says Jacob.
The researchers tested their approach by performing three different modeling tasks: inferring navigation goals from previous routes, guessing a person’s communication intentions based on their verbal cues, and predicting next moves in chess matches between people.
In every experiment, their method equaled or outperformed the popular alternative. Moreover, the researchers observed that their model of human behavior fits well with indicators of player skill (in chess matches) and task difficulty.
In the future, researchers want to apply this approach to model the planning process in other fields, such as reinforcement learning (a trial-and-error method commonly used in robotics). In the long term, they intend to continue this work to achieve the broader 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.