2020 International Conference on Social Computing, Behavioral-Cultural Modeling, & Prediction and Behavior Representation in Modeling and Simulation
Oct 19-22, 2020, Virtual

KEYNOTE SPEAKER Barry G. Silverman, PhD
Personal website: https://www.seas.upenn.edu/~barryg/
Professional Title: Professor Emeritus, Electrical and Systems Engineering, University of Pennsylvania;
Talk Topic: StateSim: A Country Modeling and Systems Engineering Journey

StateSim: A Country Modeling and Systems Engineering Journey

This is a story of perseverance in country modeling through numerous failures peppered with an increasing number of successes over time. It began 12 years ago with a DARPA “horse race” challenge to model the likely behavior of every country of the world. The winners used statistical modeling (lagged regression) and extrapolated trends in black-box fashion. With StateSim, we wanted to use causal agent-based models so one could study counterfactuals and different courses of action (COAs). We succeeded in hiring country desk experts to help us build models of four countries which passed rich validation tests. However, we failed to do all countries since each country model took several months to produce, making global coverage unsustainable. We have addressed this challenge in two ways. First, in the ensuing years we built 15 more country models and converted them into a “repository” of validated, reusable country modeling components (different types of leaders, followers, factions, institutions, resources, etc.). Second, we have exploited the rise of “Big Data,” and the dozens of open social science data sources that cover many of the required parameters. Merging our repository of reusable parts with machine learning to scrape parameters from the open data sources created a needed breakthrough. We have spent the past two years solidifying this breakthrough into a Rapid StateSim Model Generator. In trials in the USA, England, and Australia, Information Officers used the StateSim Generator to model various countries around the world and run them for numerous COA studies (in a 2-3 day window per country). Of course, the validity of each country model’s projections vary as a function of (1) its coverage in the open source data, (2) the validity of what is in the open source data, and (3) the extent that the user tests and fine-tunes the model to ensure it corresponds with reality. In sum, when statistical models predict a conflict, instability, or other scenario of interest, one must turn to causal models to study the “what-ifs” – direct, secondary, and tertiary impacts of COAs and policy choices. Sufficient progress (coverage, speed, ease of use, validity, etc) has occurred so that it is now becoming reasonably fast to build and run causal models of any country’s instability issues.