To build ai systems that can collaborate effectively with humans, it helps to have a good model of human behavior to start with. But humans tend to behave suboptimally when making decisions.
This irrationality, which is especially difficult to model, often boils down to computational limitations. A human being cannot spend decades thinking about the ideal solution for a single 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 limitations that may hinder the agent's problem-solving abilities.
Their model can automatically infer an agent's computational limitations by seeing just a few traces of its previous actions. The result, the so-called “inference budget” of an agent, can be used to predict the future behavior of that agent.
In a new paper, the researchers demonstrate how their method can be used to infer someone's navigation goals from previous routes and to predict players' subsequent moves in chess games. Their technique matches or surpasses another popular method for modeling this type of decision making.
Ultimately, this work could help scientists teach ai systems how humans behave, which could allow these 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, electrical engineering and computer science (EECS) graduate student and lead author. of a article about this technique.
“If we know that a human is about to make a mistake, having seen how they have behaved before, the ai agent could intervene and offer a better way to do it. Or the agent could adapt to the weaknesses that his human collaborators have. Being able to model human behavior is an important step towards creating an ai agent that can actually help that human being,” she says.
Jacob wrote the paper with Abhishek Gupta, assistant professor at the University of Washington, and lead author Jacob Andreas, associate professor at EECS and member of the Computer Science and artificial intelligence Laboratory (CSAIL). The research will be presented at the International Conference on Learning Representations.
Modeling behavior
Researchers 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 the agent always choosing the correct option, the model could have that agent make the correct choice 95 percent of the time.
However, these methods may not capture the fact that humans do not always behave suboptimally in the same way.
Others at MIT have also studied more effective ways to plan and infer objectives in the face of suboptimal decision making.
To build their model, Jacob and his collaborators were inspired by previous studies of 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 challenging matches.
“At the end of the day, we saw that the depth of planning, or how long you think about the problem, is a very good indicator of how humans behave,” Jacob says.
They built a framework that could infer the depth of an agent's planning from previous actions and use that information to model the agent's decision-making process.
The first step of their method involves running an algorithm for a given period of time to solve the problem being studied. For example, if they are studying a game of chess, they could let the game algorithm run for a certain number of steps. In the end, researchers can see the decisions the algorithm made at each 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 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. You can use the inference budget to predict how that 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 performing any additional work. This framework could also be applied to any problem that can be solved with a particular class of algorithms.
“For me, the most surprising thing was the fact that this budget of inferences is very interpretable. That is to say that more difficult problems require more planning or that being a strong actor means planning for longer. When we first set out to do this, we didn't think our algorithm would be able to detect these behaviors naturally,” says Jacob.
The researchers tested their approach on 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-to-human chess games.
Their method matched or outperformed a popular alternative in every experiment. Furthermore, the researchers saw that his human behavior model matched well with measures of player skill (in chess games) and task difficulty.
In the future, the researchers want to use this approach to model the planning process in other domains, such as reinforcement learning (a trial-and-error method commonly used in robotics). In the long term, they intend to continue developing this work towards the broader goal of developing more effective ai collaborators.
This work was supported, in part, by the artificial intelligence for Augmentation and Productivity program at the MIT Schwarzman School of Computing and the National Science Foundation.
