The Design of Automated Validation and Explanation for Large-Scale Social Agent Systems

Socio-technical problems, such as how smallpox outbreaks would spread in and affect modern societies, often have complex interrelated parts that defy simple mathematical analyses. A promising toolkit to solve these problems is large-scale multi-agent models, whose subsets with stochastic and knowledge-intensive networked interactions are social agent models. The value of these models and their simulations increases significantly if they can effectively exploit existing data-streams and knowledge for validation and explain emergent behaviors. Most of the existing technology for validating computational models is designed for deterministic and/or small scale systems where it is often possible to obtain validation manually or semi-automatically by bruteforce. Large-scale social agent systems pose an entirely different set of challenges. Given the size of such systems, the vast quantities and variable quality of empirical data involved, automated validation and explanation approaches are crucial. In this paper, such an approach is described in the design of an automated validation and explanation tool called WIZER that utilizes knowledge-intensive simulation-aided search and inference techniques -and knowledge-based control of simulation -capable of principled exploration of the parameter and model space, constrained by empirical data and knowledge. WIZER inference engine is built upon our novel Probabilistic Argumentation Causal System, derived from Probabilistic Argumentation Systems and Causal Analysis. Contact: Alex Yahja Institute for Software Research International Carnegie Mellon University, Pittsburgh, PA 15213 Tel: 1-412-268-7527 Fax: 1-412-268-6938 email: ay@cmu.edu