Humans of Simulated New York

“I feel, personally, that the study of experimental games is the proper route of travel for finding the ‘ultimate truth’ in relation to games as played by human players.”

John Nash, The Agencies Method For Modeling Coalitions and Cooperations In Games


For their research residency at the DBRS Innovation Lab, designers Francis Tseng and Fei Liu teamed up to build an agent-based simulation of an economic model. Based on data culled from the American Community Survey, a supplemental census containing economic data. Liu and Tseng looked at New York City data from 2004–2014 to produce a model city whose operations were determined by the relationships between data points.

For any deep financial analysis of data an accurate model is needed. The consequences of a model that is incorrect or contains incorrect assumptions can be devastating. The model must be comprehensive enough to account for the most important factors that could have an impact on the loan or transaction, but elegant enough to be useful. Deciding which elements of a dataset are salient to a deal is the critical work of an economic modeler.

In their month-long residency, Liu and Tseng developed a project that asks: What kinds of assumptions make their way into models as artifacts of the technologies that are used to produce them? How can design designers communicate these assumptions in such a way as to make them available for discussion?

Their simulation is an agent-based model, meaning that it is composed of a collection of autonomous agents (“simulants” as Liu and Tseng called them) who are capable of acting independently and influencing one another with their actions. Each simulant in Liu and Tseng’s simulation is programmed to behave as though they want certain things — a job, a higher standard of living — and even exists in a social network that determines its likeliness to get a job (or a contagious disease).

Illustration of autonomous agency. For a hands-on introduction to modeling with autonomous agents, see Daniel Shiffman’s Nature of Code. (Image: Joelle Fleurantin)

DBRS Innovation Lab Director Amelia Winger-Bearskin first began to research multi-agency as a professor at Vanderbilt University. “I was interested in how multi-agency and interactivity could redefine models of large systems,” she said. “It was this area of research that first got me excited to study the art that was part of the structure of computer science.”

“Agent-based modeling is one of the most modern techniques of mathematical modeling,” write Leonid Hulianytskyi and Diana Omelianchyk. “The main idea of this approach is in conception of economics as the complex adaptive system whose behavior is defined by multiple interactions between autonomous heterogeneous economic agents that feature various behavior patterns and ability to learn.”

Agent-based models are a kind of multi-agent system, a concept that was formerly the provenance of game theorists. With the development of increasingly powerful GPU’s, theorists and researchers had the processing power to model these systems to a degree of complexity and detail that was previously unimaginable. Today, agent-based models have a long and established history of applications within fields such diverse as “computer science (including artificial intelligence, theory, and distributed systems), economics (chiefly microeconomic theory), operations research, analytic philosophy, and linguistics.”

As Christopher K. Chan writes in An Agent Based Model of a Minimal Economy, “the agent-based methodology has several advantages over the traditional approach to economic modeling.” In an agent-based model, individuals “can be modeled to behave with bounded rationality, rather than perfect rationality.” Agents do not have access to the entire scope of a simulation, but only to the parameters that directly influence them. This limited, imperfect rationality by which agents operate is a fundamental stipulation of game theory and behavioral economics, but one which is conventional economic modeling had difficulty expressing.

Agent-based models also allow for the representation of emergent properties of the system that can function as agents of the economy in their own right. Chan writes that “by modeling the behaviors of boundedly rational agents and observing the outcomes that arise endogenously from their simulated interactions, the agent-based approach can be employed to expand our understanding of the fundamental patterns observed in the real economy.”

An agent-based model, in other words, is a structure of entailments that is able to produce unforeseen outcomes as a result of its own immanent procedural logic. This made it the perfect tool for Liu and Tseng to experiment with visualizing the internal dynamics of a static dataset. “We took something that was flat and tried to extract points from it that could be leveraged to demonstrate a relationship,” Liu said. “The challenge was picking the more interesting points. but that’s also skewing data, that happens in data visualization all the time.”

Key representing the different characteristics of agents within the simulated city (Image: Joelle Fleurantin)

“Confirmation bias is always a problem in model building,” added Tseng. “We have to bake in assumptions about the world, but they might not be correct. We tried to get around it by opening up as many parameters as possible.”

The simulation is designed as a game played by three players. At the beginning of the simulation the users are invited to establish the baseline conditions in which the city will exist. Will food be nutritious and widely available? Is there a virulent disease poised to wreak havoc on the city’s inhabitants?

“The exogenous variables are determined by the user at the outset,” explained Tseng. “How much labor will be required to produce a unit of food, ratios that determine units of measure — these are assumptions that influence the way the model functions, and we wanted to give users control over that aspect of the simulation.”

Key representing the different structures of the simulated city (image: Joelle Fleurantin)

Once the initial terms of the simulation have been set, the city is populated by a crush of multicolored dots that swarm the streets (Simulated New York is laid out on a perfect grid) going about their simulated days, working, paying taxes, and sometimes getting sick and dying.

Users are periodically prompted to make decisions that affect the outcome of the city as a whole, raising taxes and reallocating funds. The goal is to preserve the simulated city as long as possible.

Liu and Tseng’s model is not only a model, but also contains with in it a model for the act of constructing models.

Tseng describes the decision to structure their simulation as an interactive game as “a proxy for democratic governance,” a turn of phrase which generates an productive semantic loop. On the one hand, democratic governance is itself a kind of model of human interaction that channels and codifies discourse to become action. At the same time, these models themselves are often used in government (the American Community Survey states on its website that it “helps local officials, community leaders, and businesses understand the changes taking place in their communities.”) Liu and Tseng’s model is not only a model, but also contains with in it a model for the act of constructing models.

“My gut feeling is that the next phase of this AI renaissance will be around simulation,” Tseng wrote about the project. “AlphaGo’s recent victory over the Go world champion — a landmark in AI history — resulted from a combination of deep learning and simulation techniques, and I think we’ll see more of this kind of hybrid.”

He continues:

“Simulation is important to planning, a common task in AI. Here “planning” refers to any task that requires producing a sequence of actions (a “plan”) that leads from a starting state to goal state… To produce such plans, there needs to be a way of anticipating the outcome of certain actions. This is where simulation comes in. For example, in deciding how to get to the airport, I have to consider various scenarios — traffic could be bad at this particular hour, maybe there’s some chance the cab breaks down, and so on. Simulation is the consideration of these scenarios.
“So simulation, especially as it relates to planning, is crucial to AI’s more interesting applications, such as policy and economic simulation which seeks to understand the implications of policy decisions. Much like machine learning, planning and simulation have a long history and are already used in many different contexts, from shipping logistics to spacecraft. The residency was a great opportunity to begin exploring this space.”

Read Tseng’s comprehensive and detailed writeup here.

An early version of the city

Our team consists of engineers and mathematicians, story-tellers and data artists. We interrogate big datasets to uncover hidden trends, make animations that set beautiful geometries in motion, and train machine-learning algorithms to hew insights from raw numbers. Our tools allow us to examine the details of our economy and our world with extreme precision, and to simplify complex information accurately. We are dedicated to finding exciting new ways of helping people see the insights beyond the rating. Learn more at http://dbrslabs.com/