GeoWorld

The Difficulties of Modeling 'Agency' in Space and Time DYNAMIC DECISION POINTS M any of the temporal characteristics (and applications) explored in this column often feature change that can be modeled beforehand or analyzed after the fact—weather is a good example. No matter how much meteorologists may be disparaged, they typically use a series of climate models to predict weather based on pressure gradients, wind currents and atmospheric moisture. Within a few hours or days, many models are reasonably accurate. At close to a week out, however, models BY ERIK SHEPARD may diverge wildly. Anyone who has ever looked at a series of spaghetti models for an impending hurricane can attest to this. Signs of Interdependence Erik Shepard is director and principal of Waterbridge GeoDesign Inc.; e-mail: erik@ waterbridge.biz. 30 But the key fact behind these models, and many other forms of dynamic data, is that a model can be used for forecasting or summarizing the past vis-à-vis regression: a model based on a series of interdependent variables. When attempting to model people and their activities in the past or future, however, such models become more complex because of "agency." Each of the people being modeled is a "free agent," with the ability to make independent decisions about past or future actions. No longer can a finite set of variables define the model. Now, each independent agent must be modeled holistically. Although people have free agency in the philosophical sense, they may not always be able to apply it, and that can determine how models are constructed. For example, in a classic application of field-workforce management, the resources being routed don't have agency. A set of variables optimizing travel time, customer satisfaction, cost and other factors determine crew routing. Resources— people—aren't free to choose routes based on desire. These types of logistics applications are well understood, and they're well modeled using classical mathematics and statistics. Agency applies to modeling independent decisions by people acting independently; a classic spatial G E O W O R L D / M A Y 2 O 1 3 example is migration. Migration doesn't necessarily mean heading south for the winter (although it can). Commuting from suburbs to the downtown center is another form of migration, albeit shorter term. Each person living in the suburbs has an independent origin and destination as well as independent routes to take him or her there in the morning and back again at night. Transportation applications attempt to predict congestion based on commutes and invest appropriately in infrastructure. But doing this in the context of agency is difficult, so incorporating dynamics into a spatial model is difficult. Even storing data about past commutes is difficult; often the best that can be done is to look at traffic counts on major thoroughfares. Dealing with Agency Multi-agent simulations (MASs) create a model in which each agent is treated independently and acts independently based on some combination of factors—minimizing travel time or distance or random factors (e.g., had to have a cappuccino this morning). The interesting thing about MASs, and people with agency, is that what appears at first to be completely random isn't always when viewed at different scales. MASs aren't necessarily spatial models; an MAS for online trading may have no explicitly spatial component (although virtually all MASs traverse some sort of space, even if defined by independent variables). But many MASs have an explicitly spatial component—an MAS for migration, ecosystem predation, disaster response or many other applications will locate agents in a map space. During the last few years, there have been several efforts to integrate MAS with GIS to visualize the movement of agents in that map space. And integrating MAS with GIS will model agency as well as visualize emergent patterns in space and time—patterns that may not be obvious when reviewing raw data. Saying Farewell Agency is one of the most challenging spatiotemporal applications, and a worthy topic to conclude this column. I've enjoyed writing this column for the last few years, and I hope there have been useful nuggets of information to help readers think about how to model time—as well as space—in a GIS. Although this was purely a research topic only a few years ago, the ability—and need—to model time and space has become mainstream and essential for today's challenges and applications. Situational intelligence is an essential application and a cornerstone of geospatial capability. The ability to model—mathematically or via MAS—is a key element toward implementing situational intelligence and dynamic decision support.

• Cover