Agent_Zero and Generative Social Science
Abstract:
Agent_Zero is a formal alternative to the Rational Actor model of Economics and Game Theory. Agent_Zero's actions result from the interaction of a non-deliberative fear module, a
boundedly-rational deliberative module, and a social module. Each module is grounded in contemporary cognitive neuroscience. When multiple Agent_Zeros interact, they generate a wide range of important collective phenomena in economics, health behavior, violence, network dynamics and even jury behavior. Epstein's book Agent_Zero: Toward Neurocognitive Foundations for Generative Social Science was published by Princeton University Press in 2014.
Bio:
Joshua Epstein is Professor of Epidemiology in the NYU School of Global Public Health, and founding Director of the NYU Agent-Based Modeling Laboratory, with affiliated appointments at The Courant Institute of Mathematical Sciences, and the College of Arts & Sciences. Prior to joining NYU, he was Professor of Emergency Medicine at Johns Hopkins, and Director of the Center for Advanced Modeling in the Social, Behavior, and Health Sciences, with Joint appointments in Economics, Applied Mathematics, International Health, and Biostatistics. Before that, he was Senior Fellow in Economic Studies at the Brookings Institution and Director of the Center on Social and Economic Dynamics. His research interest has been modeling complex social dynamics using mathematical and computational methods, notably the method of Agent-Based Modeling in which he is a recognized pioneer. For this transformative innovation, he was awarded the NIH Director’s Pioneer Award in 2008, an Honorary Doctorate of Science from Amherst College in 2010, and was elected to the Society of Sigma XI in 2018. He has applied this method to the study of infectious diseases (e.g., Ebola, pandemic influenza, and smallpox), vector-borne diseases (e.g., zika), urban disaster preparedness, contagious violence, the evolution of norms, economic dynamics, computational archaeology, and the emergence of social classes, among many other topics. His books include Nonlinear Dynamics, Mathematical Biology, and Social Science (Wiley 1997), Generative Social Science: Studies in Agent-Based Computational Modeling (Princeton, 2006), Agent_Zero: Toward Neurocognitive Foundations for Generative Social Science (Princeton, 2013), and with Robert Axtell, Growing Artificial Societies: Social Science from the Bottom Up (MIT, 1996). Dr. Epstein earned his BA from Amherst College and his Ph.D. from The Massachusetts Institute of Technology.
Summary:
What’s a sufficient explanation for social behaviors?
Typically: assume rational agents and compute Nash equilibrium solution
Proposal:
Look at cognitively plausible agents
Simulate them in an environment
Try to reproduce observed behaviors
Agent Zero
Affective, deliberative and social modules
Agents interact and affect each other’s modules
Modules will be improved over time and the current goal is to set up an initial framework with all the key components
A formal, mathematical and computational alternative to the rational actor model
Example:
Agent Zero’s environment interpreted as exposure to violence
2D grid of sites through which agents move
Sites can “attack” agents and stimulate them to become afraid
Cause fear conditioning that lasts for a long time after the initial stimulus
Each agent’s solo disposition depends on their private history
Agents are also affected by dispositions of other agents
From these foundations we can create realistic behaviors
Agents A and B are attacked and become afraid
Agent C has never been attacked but us affected by the fear of A and B, and starts attacking others
Framework applicable to many real-world situations
Free snake
Eat food
Join lynch mob
Refuse vaccines
Etc.
Equations: minimal sufficient model that captures human behaviors
Solo disposition to act(t) = Affective(t) + Deliberative(t)
Total disposition to act(t) = Solo(t) + Sum of everyone else's Solo(t)*weight
Social contagion is based on sharing dispositions, not imitating actions
Act if Total disposition to act(t) > threshold
Fear equation
Rescola-Wagner model
Fear acquisition
Fear extinction
Modeled based on the neuroscience of the amygdala circuit
Associative fear conditioning: a light alone doesn’t make you afraid but if you consistently shock a person while showing a light then after a while the light alone will cause fear
Fear contagion
Various neurological mechanisms (e.g. mirror neurons)
Experiments show that if a person watches another person undergoing fear conditioning training, they’ll become afraid when they see the light.
The same fear circuitry is routinely adapted to novel scenarios
Arab Face + 9/11 + fear
Doctor + Tuskegee + distrust
Financial asset + devaluation + panic
Fear is quickly learned and slowly forgotten over time
Exponential decay
Moving window average
Formula: Weight * Initial fear stimulus * exponential decay + memory (moving window average of stimuli)
Weight = the surprise level and salience level (how meaningful it is) of stimulus
Agent Zero exhibits some standard biases (biased by local and social influences)
There are many others
Anchoring bias
Loss aversion
…
Parameters
Fear learning rate
Memory window
Threshold (equal across weights)
Must define stimulus landscape and initial conditions
Network effects
Can create networks exogenous
Endogenous models
Connect agents that are similar in attribute space
Social influence proportional to agent similarity
Example: Jury dynamics
Pre-trial: no communication among agents before trial
During trial: lawyers influence jurors but jurors still don’t communicate
Jury deliberations: jurors interact and can acquit or convict due to social interactions even if none of them would do this on their own
Extension example: COVID fear
People are afraid due to COVID
They hide for a while, not afraid
Go back out, increase infections
Get afraid again, hide
Fear of disease can be combined with fear of vaccines, which can inflame disease and its fear.
Model produces multiple waves of infection, with secondary waves larger than the first one.
Extension example: addiction
Early on in drug us: deliberative module dominates
Later on: affective module most important