To create AI systems that can collaborate effectively with humans, it helps to have a good model of human behavior from the start. But humans tend to behave suboptimal when making decisions.
This irrationality, particularly difficult to model, often comes down to computer constraints. A human being cannot spend decades thinking about the ideal solution to a given 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 accounts for unknown computational constraints that may hamper the problem-solving abilities of the agent.
Their model can automatically infer an agent's computational constraints by seeing just a few traces of its previous actions. The result, called the “inference budget”, can be used to predict the future behavior of this agent.
In a new paper, the researchers demonstrate how their method can be used to infer a person's navigation goals from previous routes and to predict players' subsequent movements in chess matches. Their technique matches or outperforms another popular method of modeling this type of decision-making.
Ultimately, this work could help scientists teach AI systems how humans behave, which could allow those systems to be more responsive to their human collaborators. Being able to understand a human's behavior, and then infer their goals from that behavior, could make an AI assistant much more useful, says Athul Paul Jacob, a graduate student in electrical and computer engineering (EECS) and lead author of a reference work. article on this technique.
“If we know a human is about to make a mistake, after seeing how they behaved before, the AI agent could step in and suggest a better way to do it. Or the agent could adapt to the weaknesses of its human collaborators. Being able to model human behavior is an important step towards creating an AI agent that can actually help that human,” he says.
Jacob wrote the paper with Abhishek Gupta, an assistant professor at the University of Washington, and lead author Jacob Andreas, an associate professor at EECS and a member of the Computer Science and Artificial Intelligence Laboratory (CSAIL). The research will be presented at the International Conference on Representations of Learning.
Modeling behavior
Researchers have been building computer models of human behavior for decades. Many prior approaches attempt to account for suboptimal decision making by adding noise to the model. Instead of the agent always choosing the right option, the model could ask the agent to make the right choice 95% of the time.
However, these methods may not take into account that humans do not always understand behave suboptimal in the same way.
Others at MIT have also studied more effective ways planning and deriving goals in the face of suboptimal decision-making.
To build their model, Jacob and his collaborators drew inspiration from previous studies on chess players. They noticed that players took less time to think before acting when making simple moves and that stronger players tended to spend more time planning than weaker ones in difficult matches.
“Ultimately, we found that depth of planning, or how long someone thinks about the problem, is a very good indicator of how humans behave,” Jacob says.
They built a framework for inferring the depth of an agent's planning from prior actions and using this information to model the agent's decision-making process.
The first step of their method consists of running an algorithm for a defined duration to solve the problem studied. For example, if they are studying a game of chess, they can let the chess game's algorithm run for a certain number of steps. At the end, researchers can see the decisions the algorithm made at each step.
Their model compares these decisions to the behaviors of an agent solving the same problem. It will align the agent's decisions with those of the algorithm and identify the step at which the agent stopped planning.
From there, the model can determine the agent's inference budget, or how long that agent will plan for this problem. It can use the inference budget to predict how this agent would react when solving a similar problem.
An interpretable solution
This method can be very effective because researchers can access the full set of decisions made by the problem-solving algorithm without doing any additional work. This framework could also be applied to any problem that could be solved with a particular class of algorithms.
“For me, the most striking thing was the fact that this inference budget is very interpretable. This is like saying that harder problems require more planning or that being a strong actor means planning for a longer time frame. When we decided to do this, we didn’t think our algorithm would be able to detect these behaviors naturally,” explains Jacob.
The researchers tested their approach in three different modeling tasks: inferring navigation goals from previous routes, guessing a person's communicative intent from their verbal cues, and predicting subsequent movements during game matches. failures between humans.
Their method matched or outperformed a popular alternative in each experiment. Additionally, the researchers found that their model of human behavior matched well with measures of player skill (in chess games) and task difficulty.
In the future, researchers want to use this approach to model the planning process in other areas, such as reinforcement learning (a trial-and-error method commonly used in robotics). In the long term, they intend to continue building on 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 by the National Science Foundation.