pedestrians at sunset near the arc de triomphe

Project Details

Contacts
Samarth Swarup
Funding Agency

National Science Foundation

The modeling of human behavior is a powerful tool that can be used by policy makers to direct emergency operations, economic development and more.

The methods and approaches for the alignment of models of human behavior--coming from social psychology, cognitive science, sociology and economics--with simulation approaches that model networks within populations of humans are underdeveloped. Yet, by all accounts, some social systems and some phenomena directly implicate models of human behavior simultaneously with networked population dynamics, e.g., COVID-19. Ignoring such implications is possible, but will not advance the scientific basis of social simulation that should eventually serve as the foundation for policy and preparedness needs of governments and communities alike across a variety of domains (e.g., information operations, epidemiological surveillance and decision making, etc.).

 

Project Overview

A central challenge is the development of theoretical constructs concerning human behavior that are implementable in population-level, networked simulations. We need methods that begin to integrate networked, social observations with well-defined psychological methods—essentially to build models of psychological phenomena in the context of realistic social behavior. 

Our work attempted to define and create these types of models and constructs, which can be used to calibrate social simulations.

Findings
  • Psychological parameters of a model can affect the outcomes on a social network, e.g., Systems of Behavior and Population Health.
  • Sensitive reaction time measures are possible in uncontrolled, self-directed settings.
  • We conducted experiments where human subjects played the Iterated Prisoners’ Dilemma (IPD) against bots playing various fixed strategies (unknown to the human subjects). Results showed that prior models of human behavior in the IPD are incomplete. We presented a new model, Majority Wins, that provides a better fit to the data.

 

Figures

Human-level computational models of attitude formation
Figure 1. Human-level computational models of attitude formation are implemented, largely for theoretical but also historic reasons, using some form of a fully recurrent neural network. Panel A illustrates one such model under development; the graphical structure illustrates clustering by valence of the attitude features (positive/negative features); Panel B shows the model’s respective state space attractors. One issue with such models, in respect to their usefulness for social simulation, is a paradigm for development that incorporates more naturalistic, real-world measurements in situ. Panel C shows results towards methodological paradigms for measuring the degree of automaticity associated with features of an attitude object (e.g., the aspects/features of a presidential candidate that drive disposition toward or away from said candidate). Measurements of this class are highly sensitive to the experimental conditions/context of the experiment; they require sustained attention, complex instruction, etc. Little is known respecting how to make such measurements outside of a highly controlled laboratory environment. Our approach shows promise for the development of paradigms that can successfully implement complex psychological measurements outside of the laboratory. Panel C shows similar distributions of reaction times for attitude judgments both in situ (an uncontrolled, online portal) and in a laboratory setting. Panel. D illustrates the degree to which the nodes of a network become sorted on an attitude attribute (i.e., the attitude state of nodes is ‘correlated’ locally more so than globally with respect to the networks topology) as a function of two parameters in a psychological, agent-level model of attitude; this is a demonstration with synthetic, stylized data and model of the implications of the psychological level of scale with the sociological, more aggregate level of scale.
Tools

Multi-agent neural network simulator

C Based Multi-agent neural network simulator