Researchers from MIT and the University of Washington have developed a model to predict the behavior of human and artificial intelligence (AI) agents, taking into account computational constraints. The model automatically deduces these constraints by processing previous actions of the agent. This “inference budget” can help predict future behavior of the agent; for instance, it can be used to infer navigation goals from past routes or predict upcoming moves in a chess game.
This work could enable AI systems to better understand human behavior to improve collaboration. For example, an AI assistant able to recognise a user’s behavior and infer his or her goals can offer timely and more efficient guidance. “If we know that a human is about to make a mistake…the AI agent could step in and offer a better way to do it,” says lead author Athul Paul Jacob from EECS.
The model was inspired by studies of chess players who strategize more deeply in complex matches. Researchers developed an algorithm that could infer an agent’s depth of planning from previous actions and use this information to predict their decision-making. By comparing the agent’s decisions to the problem-solving algorithm, it could be determined where the agent ceased planning. The agent’s “inference budget” or the estimation of how much planning time it would allocate to a particular problem can then be established.
The simplicity of the inference budget was a striking feature, according to Jacob, who pointed out that it could determine the need for longer planning times for difficult problems or for stronger players. The method was tested in three different scenarios: inferring navigation goals from prior routes, guessing communicative intent from verbal cues, and predicting subsequent moves in chess matches. The model performed as well as, or even better than, current popular alternatives.
In the future, researchers plan to apply this method to model the planning process in other domains like reinforcement learning, a technique frequently used in robotics. The ultimate goal is to create more effective AI collaborators. Funding for this project was partly provided by the MIT Schwarzman College of Computing Artificial Intelligence for Augmentation and Productivity program and the National Science Foundation.
