To construct AI systems that may collaborate effectively with humans, it helps to have a superb model of human behavior to start out with. But humans are inclined to behave suboptimally when making decisions.
This irrationality, which is very difficult to model, often boils right down to computational constraints. A human can’t spend many years desirous about the perfect solution to a single problem.
Researchers at MIT and the University of Washington developed a technique to model the behavior of an agent, whether human or machine, that accounts for the unknown computational constraints which will hamper the agent’s problem-solving abilities.
Their model can mechanically infer an agent’s computational constraints by seeing just a number of traces of their previous actions. The result, an agent’s so-called “inference budget,” will be used to predict that agent’s future behavior.
In a brand new paper, the researchers reveal how their method will be used to infer someone’s navigation goals from prior routes and to predict players’ subsequent moves in chess matches. Their technique matches or outperforms one other popular method for modeling one of these decision-making.
Ultimately, this work could help scientists teach AI systems how humans behave, which could enable these systems to reply higher to their human collaborators. With the ability to understand a human’s behavior, after which to infer their goals from that behavior, could make an AI assistant far more useful, says Athul Paul Jacob, an electrical engineering and computer science (EECS) graduate student and lead writer of a paper on this method.
“If we all know that a human is about to make a mistake, having seen how they’ve behaved before, the AI agent could step in and offer a greater technique to do it. Or the agent could adapt to the weaknesses that its human collaborators have. With the ability to model human behavior is a crucial step toward constructing an AI agent that may actually help that human,” he says.
Jacob wrote the paper with Abhishek Gupta, assistant professor on the University of Washington, and senior writer Jacob Andreas, associate professor in EECS and a member of the Computer Science and Artificial Intelligence Laboratory (CSAIL). The research shall be presented on the International Conference on Learning Representations.
Modeling behavior
Researchers have been constructing computational models of human behavior for many years. Many prior approaches attempt to account for suboptimal decision-making by adding noise to the model. As a substitute of the agent all the time selecting the right option, the model may need that agent make the right selection 95 percent of the time.
Nonetheless, these methods can fail to capture the proven fact that humans don’t alwaysbehave suboptimally in the identical way.
Others at MIT have also studied simpler ways to plan and infer goals within the face of suboptimal decision-making.
To construct their model, Jacob and his collaborators drew inspiration from prior studies of chess players. They noticed that players took less time to think before acting when making easy moves and that stronger players tended to spend more time planning than weaker ones in difficult matches.
“At the tip of the day, we saw that the depth of the planning, or how long someone thinks concerning the problem, is a very good proxy of how humans behave,” Jacob says.
They built a framework that might infer an agent’s depth of planning from prior actions and use that information to model the agent’s decision-making process.
Step one of their method involves running an algorithm for a set period of time to unravel the issue being studied. As an example, in the event that they are studying a chess match, they could let the chess-playing algorithm run for a certain variety of steps. At the tip, the researchers can see the choices the algorithm made at each step.
Their model compares these decisions to the behaviors of an agent solving the identical problem. It’s going to align the agent’s decisions with the algorithm’s decisions and discover the step where the agent stopped planning.
From this, the model can determine the agent’s inference budget, or how long that agent will plan for this problem. It might use the inference budget to predict how that agent would react when solving an analogous problem.
An interpretable solution
This method will be very efficient since the researchers can access the complete set of selections made by the problem-solving algorithm without doing any extra work. This framework is also applied to any problem that will be solved with a selected class of algorithms.
“For me, probably the most striking thing was the proven fact that this inference budget may be very interpretable. It’s saying tougher problems require more planning or being a robust player means planning for longer. After we first got down to do that, we didn’t think that our algorithm would have the opportunity to select up on those behaviors naturally,” Jacob says.
The researchers tested their approach in three different modeling tasks: inferring navigation goals from previous routes, guessing someone’s communicative intent from their verbal cues, and predicting subsequent moves in human-human chess matches.
Their method either matched or outperformed a well-liked alternative in each experiment. Furthermore, the researchers saw that their model of human behavior matched up well with measures of player skill (in chess matches) and task difficulty.
Moving forward, the researchers wish to use this approach to model the planning process in other domains, corresponding to reinforcement learning (a trial-and-error method commonly utilized in robotics). In the long term, they intend to maintain constructing on this work toward the larger goal of developing simpler AI collaborators.
This work was supported, partly, by the MIT Schwarzman College of Computing Artificial Intelligence for Augmentation and Productivity program and the National Science Foundation.
