Researchers at MIT and the University of Washington have created a model to predict the decision-making behavior of both human and AI agents, even in the presence of unknown computational constraints. The system is designed to infer the ‘inference budget’ of a given agent, in other words, how much time or computational resource they are likely to dedicate to a specific problem. It does this by studying traces of the agent’s previous actions and then uses this to predict future behavior.
The research team, led by Athul Paul Jacob, demonstrated their model by using it to infer patterns of navigation and predict subsequent moves in chess games from previous actions. They found that it either matched or outgrew a popular method used to model decisions. This suggests potentially significant benefits in helping teach AI systems to work more effectively with human collaborators.
To create the model, the team took inspiration from studies of chess players, who appeared to take less time to think before executing simpler moves and longer times for more complex ones. Comparing AI and human decisions in simulated tasks allowed the researchers to identify at what point the human stopped planning and determine their ‘inference budget’.
Moreover, the computational efficiency of the new method is high, as researchers can access the entirety of the decision-making process run by an algorithm without any additional work. Jacob and his teammates were pleased to discover that the technique naturally demonstrated a strong understanding of the behaviors observed in agents.
The model was evaluated against three separate tasks: inferring navigation goals from previous routes taken, guessing communicative intent from verbal cues, and predicting subsequent moves in a human chess match. The model displayed a strong alignment with measures of player skill – in the chess task – and task difficulty.
The research, partly funded by the MIT Schwarzman College of Computing Artificial Intelligence for Augmentation and Productivity program and the National Science Foundation, is set to be presented at the International Conference on Learning Representations. The team aims to apply their model to other domains like reinforcement learning and continue enhancing it for more effective AI collaboration.
