The plethora of data available through social media and the internet of things has led to many recent successes in predictive modeling of societal outcomes. Yet, these models often fail to provide human-interpretable explanations of the mechanisms of human decision-making that generate these statistical relationships. Demystifying the nature of these mechanisms is fundamental to our understanding of human behavior and for the design of effective policies and interventions that may benefit society.
In this talk, I discuss the challenges associated with extracting human-interpretable, individual-scale explanations of complex social phenomena and introduce a novel, explainable artificial intelligence framework, Evolutionary Model Discovery, which automates the inference of mechanistic explanations through artificial societies. This framework quantifies the importance of hypothesized factors of human-decision making on populations of genetically-programmed agent-based models through random forest feature importance evaluation. Evolutionary Model Discovery assisted in the successful causal inference of three very different cases of complex human social behavior: 1) previously unconsidered factors driving the socio-agricultural behavior of an ancient, ancestral Pueblo civilization are discovered, constructing a more robust and accurate version of the Artificial Anasazi model; 2) factors leading to the coexistence of mixed patterns of segregation and integration are discovered on a recent extension of Schelling's Segregation model; 3) factors determining response prioritization by social media users under information overload are discovered on an ensemble of a model of information overload and the Multi-Action Cascade Model of conversation.
Chathika Gunaratne received his Ph.D. in Modeling and Simulation from the University of Central Florida in 2019. He holds a M.S. in Modeling and Simulation also from UCF and a B.Sc. in Computer Science from the University Of Colombo, Sri Lanka. Chathika’s research involves the development of data-driven modeling and simulation techniques for the investigation of complex social behavior. In his dissertation work, he developed an explainable artificial intelligence framework, Evolutionary Model Discovery, and its open-source software toolkit for factor importance analysis of genetically programmed agent-based models. Evolutionary Model Discovery has been successfully applied to infer the generating mechanisms of a variety of complex social systems, from the socio-agricultural behavior of an ancient society to response prioritization under information overload on online social media. He is also currently engaged in investigating information dynamics and narrative evolution through at-scale, GPU-based models of information diffusion. Chathika’s research has been supported by DARPA, AWS, and Microsoft Azure grants. He has also worked in the simulation industry both as a serious-games developer, developing crowd pattern-of-life for SimCentric Technologies, and as a modeling and simulation specialist for Universal Studios, using agent-based modeling to optimize guest movement at the Orlando theme park.