Researchers from MIT and the University of Washington have developed a method to model the behaviour of an agent, including its computational limitations, predicting future behaviours by examining prior actions. The method applies to both humans and AI, and has a wide range of potential applications, including predicting navigation goals from past routes and forecasting future moves in chess matches. The technique could surpass existing methods for modelling decision-making processes.
Understanding and predicting human behaviour is crucial in Artificial Intelligence (AI) development, as this understanding aids AI’s ability to work more effectively with humans. By predicting a human’s likely mistakes based on their past behaviour, an AI assistant could intervene to suggest more effective solutions or adjust to the human’s weaknesses. This has significant implications for the utility and value of such AI assistants, as it could expand their practical applicability in a range of contexts.
A common method for modelling human behaviour is to incorporate a degree of “noise” into the model, reflecting the volatility of human decision-making. This usually involves programming the AI agent to make the correct decision most of the time, replicating the human capability of making the right decision, but not always. However, those methods overlook the variability in human behaviour; humans do not always behave suboptimally in predictable or consistent ways. The new method pursued by the MIT and the University of Washington tries to resolve these shortcomings.
This new method draws on chess players’ behaviour, where the time to think before acting and the amount of planning involved varies according to complexity of moves and expertise of players. The research team utilised this data to construct a model of an agent’s decision-making process, leading to the development of an interpretable solution. This solution reveals that more complex problems need more extensive planning, and strength in players usually correlates with longer planning times.
The method was tested on three different modeling tasks in addition to its original experiments. Here, it either matched or outperformed the rival technique. For example, in a chess match, the research team saw a strong correlation between their model of human behaviour and the player’s expertise and difficulty of the task.
Moving forward, the team is planning to apply this method to other domains like reinforcement learning, a trial-and-error methodology used efficiently in robotics. They aim to continue developing more effective AI collaborators that can help humans and provide assistance in more complex tasks. The research project was partly funded by the MIT Schwarzman College of Computing Artificial Intelligence for Augmentation and Productivity program and the National Science Foundation.
