CS Colloquium: David Pynadath (USC ICT) - Data-Driven Modeling of Human Social Behavior with Recursive Decision-Theoretic Agents
Tue, Jan 21, 2020 @ 11:00 AM - 12:00 PM
Conferences, Lectures, & Seminars
Speaker: David Pynadath, USC / ICT
Talk Title: Data-Driven Modeling of Human Social Behavior with Recursive Decision-Theoretic Agents
Abstract: Social scientists, policy makers, and other analysts have increasingly turned to multiagent social simulation as a generative methodology for representing, analyzing, and simulating human behavior. Typical agent-based social simulation methods are attractive, because they use simple, reactive rules that are directly expressible by the people seeking to use them. In contrast, AI provides algorithms for generating autonomous decisions that can match a human level of complexity, but that same complexity is a currently insurmountable obstacle to their use by AI non-experts.
At ICT, we have developed a social simulation framework, PsychSim, using decision-theoretic agents with a theory of mind (ToM) to form mental models about others and use those models to inform their own decision-making. While PsychSim's recursive Partially Observable Markov Decision Processes (POMDPs) offer a generative and transparent approach to social simulation, they share the disadvantage of similarly complex AI languages in that much effort and, often, much error is incurred when building models in them. Fortunately, the growing availability of data about people, their perceptions, and their behaviors offers a novel opportunity for automated support to both reduce the burden and increase the accuracy of the modeling process.
In this talk, I will present algorithms we have developed and applied to two different scenarios: (1) response of an urban population to a disaster, and (2) perceptions of inequality among different national, ethnic, and religious populations. In particular, we analyze the results of applying different automated methods for identifying dynamic influence diagrams whose output matches the beliefs and behaviors that people exhibit in these two scenarios. Because no single model correctly predicted everyone's perceptions and behaviors, we had our algorithm select additional models to capture atypical cases as well. Even with a very restricted space of candidate graphs, our algorithms found multiple models consistent with many of the people in the data sets. We quantify the ambiguity in the models selected by analyzing these cases, and, because of the graphical representation, we can compare models against each other to characterize potential differences in perceptions and behaviors. The result is an automated process that not only generates models for use within multiagent social simulation, but also quantifies the degree of confidence one can place in those models.
This lecture satisfies requirements for CSCI 591: Research Colloquium
Biography: Dr. David Pynadath is the Director for Social Simulation Research at USC ICT. He received his Ph.D. from the University of Michigan, Ann Arbor in 1999. He has published papers on multiagent systems, teamwork, social simulation, human-robot interaction, explainable AI, and plan recognition. He is the co-creator and maintainer of PsychSim, the multiagent social simulation framework that was the foundation of the work to be presented. Dr. Pynadath has collaborated with partners in academia and government to apply PsychSim to drive virtual characters in interactive simulations for teaching urban stabilization operations, cross-cultural negotiation, disaster response, and avoiding risky behavior.
Host: Jon May
Audiences: Everyone Is Invited
Contact: Cherie Carter