Researchers at MIT and the University of Washington have developed a computational model that can predict an intelligent agent’s behaviors based on its “inference budget” (i.e. the limits on its computational resources). This was accomplished by using an algorithm that recorded all the decisions made by the agent within a given period of time. They found that this model could even reveal the depth of an agent’s planning based on their actions. This technique could enable AI systems to understand and anticipate human behaviors, and so improve virtual or robotic collaborators.
People often make suboptimal decisions because they have limited time or resources to find the best solution. This phenomenon is especially hard to predict and model, but it motivated researchers to search for a better understanding of computational constraints in intelligent agents. The researchers found that an agent’s planning depth, or how long they would think about a problem before deciding, acted as a good predictor for their future behavior.
To build this model, the researchers drew inspiration from studies of chess players, who tended to think less before simple moves than before complex ones and who varied their planning based on the difficulty of the game. The researchers built a framework which could infer an agent’s planning depth from its past actions and then use that data to predict and model future decision-making processes.
The first step in their methodology involved running an algorithm for a set length of time to solve a specific problem, recording all the decisions made along the way. These decisions were then compared to those made by an agent trying to solve the same problem. By matching the two sets of decisions, the researchers could determine at which point the agent stopped planning and infer its “inference budget” (how long an agent would plan for a given problem). Furthermore, the system could use this inference budget to predict how the agent would behave when solving similar problems in the future.
The researchers tested this method on predicting navigation goals from past routes, predicting the meaning behind verbal cues, and guessing the next move in chess games between people. In all three cases, their technique either matched or outperformed a more traditional model. They found that their model’s predictions correlated well with measures of task complexity and the skill levels of chess players.
In the future, the researchers hope to apply this method to modeling planning processes in domains like reinforcement learning, a type of machine learning algorithm used in robotics. If they are successful, their technique could soon help build AI systems that understand, predict, and respond more effectively to human behavior. This research was, in part, supported by the MIT Schwarzman College of Computing’s AI for Augmentation and Productivity program and the National Science Foundation.
