MIT and University of Washington researchers have created a method to model both human and machine behaviours, taking into account unknown computational constraints which can limited problem-solving skills. The model infers an “inference budget” from previous actions. The inference budget can then predict the agent’s future behaviour. Their technique can be used to predict navigation goals and subsequent moves in a chess game and has been shown to outperform other popular methods for modeling decision-making.
According to lead author Athul Paul Jacob, this could teach AI systems how humans behave and respond, making AI assistants more optimal. If a human is about to make an error, the AI agent could intervene and offer a better solution. The ultimate goal is for the AI agent to be able to model human behaviour and assist the human more effectively.
Computational models of human behaviour have been under development for years, often trying to account for suboptimal decision-making by adding noise to the model so the agent does not always make the correct choice. This method does not always capture the fact that humans do not consistently behave suboptimally.
Jacob and his associates noticed that chess players took less time to think before making simple moves, and stronger players spent more time strategising than weaker players in complex matches. With this inspiration, they built a framework that could account for an agent’s depth of planning from prior actions and use this information to model the decision-making process.
The model was highly effective because researchers could access the full set of decisions made by the problem-solving algorithm without needing to do any extra research. The new framework could be applied to any problem that requires a specific class of algorithms.
The so-called inference budget is highly significant – it clarified that tougher problems require more planning and stronger players will plan for more extended periods. They put their theory to the test using three modelling tasks. The results either matched or outperformed a popular alternative in each experiment.
In the future, they aim to use this approach to model the planning process. Ultimately, the plan is to keep working on this model with the bigger aim to develop more effective AI systems. This project was partly supported by the MIT Schwarzman College of Computing Artificial Intelligence for Augmentation and Productivity program and the National Science Foundation.
