Researchers from MIT and University of Washington have developed a novel method that utilizes a good model of human behaviour, specifically involving the computational constraints in decision-making, in order to improve the collaboration between AI and humans. The unique technique of their new model permits an automatic inference regarding an agent’s computational constraints solely based on a few traces of their past activities. For instance, the combination of a chess player’s simple moves within a short period of time and extended planning time in challenging matches painted a picture of how humans behave.
The model essentially predicts behaviours while considering time spent planning by the agent, whether human or machine. The researchers implemented an algorithm for a specific timeframe to solve a problem and compared the results against the decisions made by the agent. This comparison of the agent’s decision-making process with the algorithm enabled the identification of the point where the agent’s planning halted. Consequently, the model inferred the agent’s “inference budget”—how long an agent would plan on a similar problem—and used this inference to predict the agent’s future action.
This model yielded very efficient results since complete decision sets could be extracted from the problem-solving algorithm without the necessity of extra work. It showed flexibility in its application, as it could be adopted with any problem that can be addressed with a specific class of algorithms. However, what the researchers found to be the most impressive aspect, was the model’s interpretability.
Using their model, the researchers were able to predict navigation goals from prior routes, decipher verbal cues of communicative intent, and predict subsequent moves in human-human chess matches. In each case, the model matched or surpassed the results of other popular models.
With such compelling results, the researchers are aiming to expand their approach to other domains, such as reinforcement learning commonly used in robotics. They are confident that the utilization of their model paves the roadmap for the development of much more effective AI collaborators in the future.