AI Seminar

A Unified Approach to Diverse Forms of Action Modeling and Prediction

John Laird

Computer Science & Engineering
Tuesday, March 30, 2010
4:00pm - 5:30pm
3725 Beyster Bldg. (Stained Glass Conference Room)

Add to Google Calendar

About the Event

Researchers in AI have long studied planning where an agent internally simulates possible actions to determine which action will lead to the best future situation. In AI planning systems, the knowledge to simulate the action is called the "action model", and in general it is represented within the agent as a rule-like data structure that describes the conditions under which the action can be executed, and the changes the action makes to that situation. However, recent results in cognitive neuroscience suggest that in humans the ability to imagine future situations is sometimes tied to episodic memory recall. Together these results raise the question as to what are the possible processes and sources of knowledge an agent can use to predict the future, and how those different types of knowledge are learned. Our own hypothesis is that there are many different sources of knowledge available to an agent that can be used for action modeling, but that these are embedded within a unified framework. In this talk, I describe such a framework based on the Soar cognitive architecture, where different processes and sources of knowledge are available for prediction, including rules, episodic memory, semantic memory, mental imagery, and general problem solving. A running Soar model that includes these processes will be demonstrated on a simple task. Joint work with Joseph Xu and Sam Wintermute.

Additional Information

Sponsor(s): Toyota

Open to: Public