Researchers at MIT and the University of Washington have developed a model that can predict an agent’s potential computational limitations, and therefore their decision-making process, simply by observing past behaviour. Referred to as an “inference budget,” this could enable AI systems to better predict human behaviour. The research paper demonstrates this modelling method within the context of navigating routes and predicting moves in chess games.
Understanding human behaviour and predicting based on this understanding is a key component in creating AI that can meaningfully assist human individuals. “If we know that a human is about to make a mistake, having seen how they have behaved before, the AI agent could step in and offer a better way to do it,” explains Athul Paul Jacob, author of the paper.
The team derived their model from past studies of chess players, observing that the time spent planning correlated with the difficulty of the game and the skill of the player. From this, the researchers created a framework capable of inferring an individual’s ‘depth of planning’ from their previous actions, and used this information to predict their future decision-making processes.
Their model operates by running an algorithm to solve a given problem— for example, a chess game— and compares the decisions made by the algorithm at each step to the decision-making process of the agent. This allows for an identification of when the agent stopped planning, providing insight into the ‘inference budget’ or how long the agent plans for a particular problem.
The process is efficient as it allows for a full examination of the decision-making process without extra work, and is applicable to problems resolved through a particular class of algorithm. Jacob says that the inference budget is “very interpretable,” and indicates that stronger players tend to plan for longer, and that more complex problems require more planning. This runs contrary to initial expectations, and Jacob didn’t believe the algorithm would naturally identify these behaviours.
The team tested their approach in several arenas, including predicting navigation goals, understanding communicative intent, and predicting moves in chess games. The model either matched or outperformed existing alternatives in each test, and matched player skill and task difficulty well. Future plans involve applying the model to the planning process in various other areas, including reinforcement learning in robotics, with the ultimate aim of developing more effective AI collaborators. The research was partly funded by the MIT Schwarzman College of Computing Artificial Intelligence for Augmentation and Productivity program and the National Science Foundation.