Artificial Intelligence (AI) researchers at MIT and the University of Washington have created a model that can predict a human’s decision-making behaviour by learning from their past actions. The model incorporates the understanding that humans can behave sub-optimally due to computational constraints — essentially the idea that humans can’t spend indefinitely long periods considering the optimal solution to a problem.
The researchers’ model infers this ‘inference budget’ (computational constraints) from an agent’s, human or machine, previous actions to predict future behaviour. For example, the model can deduce a person’s navigational objectives from former routes or anticipate a player’s next move in a chess match.
The technology could greatly enhance AI systems’ ability to collaborate with humans. If an AI system can understand human behaviour and anticipate their goals, it can become a more effective assistant. It could, for instance, anticipate mistakes and suggest better solutions, or adapt to human weaknesses.
The research, based on prior studies on chess players, found that the depth of planning, the time spent considering a problem, is a good indicator of human behaviour. An agent’s ‘inference budget’, or the length of planning time, can be used to predict how the agent might solve similar problems in the future.
Furthermore, the model can be used efficiently as researchers can access the entire set of decisions made by the problem-solving algorithm without extra work. The model has been tested in three scenarios — inference of navigational goals, guess communication intent, and predict subsequent moves in chess. The model performed as well as or better than a popular alternative for decision-making predictions in each of these scenarios.
Researchers’ next step is to apply this model to other domains such as reinforcement learning, used extensively in robotics. In the long run, the aim is to use this work to create more effective AI collaborators.
The research was funded by the MIT Schwarzman College of Computing Artificial Intelligence for Augmentation and Productivity program and the National Science Foundation.
