Past Workshops

From NEBL

Friday, November 2, 2007 8:30 a.m. - 5:00 p.m.
Redfield Campus, room 227

Workshop on Behavior Analysis for Artificial and Simulated Agents

It is often useful to have computer simulations of real or fictional agents. Such agents might be customers in a business training simulator, herd animals in an ecological research simulator, or monsters in a videogame. In all cases the users of the simulation will have certain expectations for what is or is not realistic behavior for the agents. In practice it has proven difficult to program robustly realistic behavior for such agents when the simulated environments are rich enough to provide opportunities for a wide variety of complex behaviors. Therefore many AI researchers have turned to machine learning techniques, which offer the prospect of training artificial agents to their target behaviors rather than programming the behaviors directly. However, the goals of simulation are different from AI researchers' traditional focus on optimization: a customer, herd animal, or monster may not behave optimally according to any objective measure of task performance. Accurate simulation instead requires capturing the style of the simulated agent's behavior, the unseen policies that guide the agent's decisions. So the question arises: When we desire accurate simulation rather than task optimization, how do we measure success?

This workshop brings together researchers from the fields of machine learning, non-linear control, and behavior analysis to examine that question. The challenge of the machine learning approach to emulating behavior will be used to motivate the question with specific examples. Then a series of tutorials, analyses, and discussions will examine the question from "internal" and "external" perspectives. From the field of non-linear control we will examine what we can say about an artificial agent's behavior by applying formal analyses to the agent's control program (the internal perspective), and from the field of behavior analysis we will examine what we can say about an artificial agent's behavior by empirical evaluation of observations (the external perspective). The goal of the workshop is to bring researchers from diverse fields together to broaden our conception of the problem and the areas where it arises, and to generate new ideas about how that extended problem can be addressed as a research question.