Simulative Inference and the Semantics of Belief

Interacting effectively with another agent, whether in competition or in cooperation, often requires reasoning about that agent's mental state. We can get clues about others' mental states by observing their behavior, or by attributing attitudes we know are common to all agents, but sometimes that's not enough. Sometimes we must draw inferences based on an understanding of how others think. An exact model of another agent's reasoning mechanism is rarely available, but when the agent is similar to us, we have a good approximation: our own reasoning ability. By temporarily assuming the attitudes we attribute to another agent, and then observing where our own reasoning process leads, we can sometimes draw further conclusions about the other agent's mental state. To draw conclusions in this way is to perform simulative reasoning.

Our work on simulative inference consists of two parts: developing a logical representation for reasoning about agents' mental states, with a formal semantics that lends itself to a treatment of simulation; and implementing extensions to Epilog, a knowledge representation and reasoning system, that demonstrate the use of simulative reasoning in natural language understanding.

This work was the thesis project of Aaron Kaplan, supervised by Len Schubert.

Publications

Aaron Kaplan's home page Research in CS at U of R