Most of the data required for UrbanSim's simulations is assembled on grid cells, each 150 by 150 meters, as shown here in the inner-city neighborhood of Queen Anne in Seattle, Wash. Each grid cell contains information on such factors as property boundaries, business establishments, real estate prices, and zoning characteristics.

23 December 2003—It's a calm late afternoon in the city, when all of a sudden a giant reptilian creature appears, crushing cars, shattering building façades, and leaving a trail of havoc as it advances along downtown streets. Those who have ever played the game “SimCity,” in which the user becomes an urban planner and has to manage the growth of a virtual metropolis, know how tempting it is to evoke Godzilla's fury to shake things up a little when nothing much interesting is happening in town.

For a group of researchers at the University of Washington in Seattle, however, simulating cities is more than just a game. They invented UrbanSim, the most sophisticated city modeling and simulation software to date—a system that could not only help people see decades into the future but also one that could play a role in settling rancorous political disputes. The program simulates urban growth and lets users test different planning scenarios, much like a real-world version of “SimCity.” You can't summon raging giant reptiles, but you can forecast the effects of deploying new highways, restricting construction over wetland areas, or doubling parking prices downtown, for example.

By the end of this year, the urban planning agency in the Salt Lake City, Utah, region will complete a series of comprehensive tests that will help it decide whether UrbanSim is suitable for use in the area. If the agency says yes to the software, that will become the system's first large-scale implementation.

The software has already been tried out by planning agencies in Honolulu (Hawaii), Eugene-Springfield (Oregon), and Houston (Texas); the Seattle region, Taipei, and Paris are beginning trials as well. People from more than 60 countries have downloaded the software, which is open source and can be freely copied from the project's Web site.

Modeling the metropolis

Cities evolve in complex and often unexpected ways—sometimes in ways that surprise even experienced planners. Build a shopping mall in one location and a traffic jam may appear at another miles away. “Ultimately, most of this stuff isn't intuitive,” says Frank Southworth, a senior staff member for R&D with the Transportation, Planning, and Policy Group at the Oak Ridge National Laboratory in Knoxville, Tenn. According to Southworth, city simulators have become essential planning tools because they provide the most effective way to forecast the likely effects of different policies and new investments.

But UrbanSim developers claim the modeling tools currently employed by many planning agencies fail to capture a good deal of this complexity of urban dynamics, especially the strong interaction between how traffic grows and where households, shops, and businesses decide to locate—in short, how transportation affects land use and vice versa. “That lack of feedback is a very significant problem,” says Paul Waddell, a professor at the University of Washington's School of Public Affairs and the director of the UrbanSim project. “Plans for multibillion-dollar transportation systems can be essentially very misguided if they overestimate their benefits.”

To address this key problem, Waddell in 1995 began to build from scratch a new modeling tool that would become UrbanSim. He was later joined by computer science professor Alan Borning, and together the pair brought to the project researchers from fields as diverse as computer science, architecture, and psychology. The UrbanSim initiative has received more than US $5 million in National Science Foundation grants and is now based at the University of Washington's newly formed Center for Urban Simulation and Policy Analysis.

Another problem Waddell wanted to correct was the coarse level of geographical detail found in older models, such as the most prevalent tool in use in the United States, DRAM/EMPAL, short for Disaggregated Residential Allocation Model/Employment Allocation Model. University of Washington researchers say UrbanSim is the first system capable of simulating the land development process at the level at which it actually occurs—the individual land parcel. Only with this level of resolution, says Waddell, can you study the effects of zoning and other public policies like the promotion of more “walkable” neighborhoods, which require planners to understand what is going on at the street level.

What's more, many of the older tools—some developed more than 40 years ago—are difficult to operate, have too many constraints, and often generate forecasts in obscure ways that only a few experts can understand. “One of the major criticisms of older models is that they are very abstract, and have often been called ‘black box' models, because no one except the modeler really knows what's going on inside,” says Waddell [see sidebar, “And You Thought Your Office was Small...”.

The UrbanSim team decided early on that their model had to be clear and explainable with representations of people, things, and actions as they exist in real-life, as opposed to the abstract variables and parameters found in black box simulators. In this sense, UrbanSim is similar to “SimCity” in that it explicitly represents a city's houses and buildings, as well as their occupants.

Four main agents interact in the virtual city: households, businesses, developers, and governments. At least in this digital incarnation, agents are all single-minded people: households decide where to live and work; businesses decide where to locate and set up their jobs; developers decide where to build houses, office buildings, and manufacturing facilities; and governments decide what development policies and investments they should apply to each part of the city. These agents, however, don't deal directly with each other; their interaction happens through one of civilization's oldest assets: land. It is the land and how it is used and transformed that ultimately determines how the urban landscape evolves.

Households, for instance, are often asking: are we happy where we live? Could we move to a bigger house in a nicer neighborhood closer to the kids' school and near that new shopping mall? They make what is called a discrete choice. A household discrete choice model, therefore, gives the probability that a given family will move based on its profile—housing costs, number of workers, income, age of members, number of children, and other characteristics—and the vacant home in consideration.

But in real-life, even people with the same characteristics make different choices. And the same person might make a different choice in two different circumstances. To account for these uncertainties, UrbanSim choice models add a random component to each individual's decision. (This same method, developed in the 1970s by economist and Nobel laureate Daniel L. McFadden, now at the University of California, Berkeley, is also used to study people's behavior when choosing telephone services, transport modes, and colleges.) In other words, in the world of UrbanSim, the gods do play dice, contrary to Einstein's dictum.

Data hungry system

The decisions the UrbanSim agents make repeat annually, so the simulator evolves the city from one year to the next over a span of, say, 30 years. At any moment, the user can “zoom in” down to grid cells of 150 by 150 meters (about the size of a suburban block) and see what's in that cell—the individual parcels of land, how many people are living and working there, what kind of housing and businesses are there, and the price of real estate.

This level of modeling detail has advantages and disadvantages. A drawback is that the number of calculations necessary to determine agents' choices can grow explosively for huge cities. In fact, the kind of simulation used in UrbanSim—microsimulation—was developed in the late 1950s and early 1960s, but was not implemented due to the lack of computing power. “There's a constant trade-off between how much detail or accuracy you want in the simulation versus keeping it computationally reasonable,” says Borning, the project's co-director.