MIT and University of Washington researchers have created a model to efficiently predict human behavior, which could potentially improve the effectiveness of AI systems working with human collaborators. Humans tend to behave suboptimally when making decisions due to computational constraints and researchers have created this model to account for these human processing limitations. The model is designed to automatically reveal the computational constraints of an agent by observing the remnants of past actions, allowing future behavior to be predicted.
Athul Paul Jacob from MIT, co-author of the paper detailing their work, explains, “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. Or the agent could adapt to the weaknesses that its human collaborators have…”
Previous computational models of human behavior have attempted to deal with suboptimal decision making by introducing noise into the model. These models, however, fail to recognize that humans do not always behave suboptimally the same way each time. The researchers’ approach was instead inspired by observing chess players, with the depth of planning proving to be a key indicator of how humans behaved.
The research team’s method involves running a problem-solving algorithm for a set period of time. For example, in a chess game, the algorithm would run for a certain number of moves. The researchers would then compare the algorithm’s decision-making pattern to the actions of a human playing the equivalent game, and identify when the human player stopped planning. From this, the model would infer the player’s ‘inference budget’ — the amount of time they spent strategizing — and use that data for future behavior prediction.
Besides its impressive efficacy, the researchers were surprised to discover that the ‘inference budget’ was an extremely interpretable way of understanding human behavior, with more complicated problems and stronger competitors requiring more planning time.
Significant advancements were made when testing this method across various modeling tasks, indicating that the model aligns well with a player’s skill level and the task’s difficulty.
In the future, researchers plan to apply this method to other areas, such as reinforcement learning in robotics, as well as further develop the model to improve AI collaboration. The project was supported in part by the MIT Schwarzman College of Computing Artificial Intelligence for Augmentation and Productivity program and the National Science Foundation.
