Agent-based models have been developed for a diverse range of subject areas, such as: archaeological reconstruction of ancient civilizations (Axtell et al., 2002; Gumerman et al., 2003) ; understanding theories of political identity and stability (Lustick, 2002); understanding processes involving national identity and state formation (Cederman, 2001); biological models of infectious diseases (Eidelson and Lustick, 2004; Chen et al., 2006); growth of bacterial colonies (Kreft et al., 1998; Krzysztof et al., 2005); single- (Emonet et al., 2005) and multi-cellular level interaction and behavior (Athale and Deisboeck, 2006); alliance formation of nations during the Second World War (Axelrod and Bennett, 1993); modeling economic processes as dynamic systems of interacting agents (Agent-based Computational Economics, ACE, Tesfatsion, 2006a); company size and growth rate distributions (Axtell, 1999); size-frequency distributions for traffic jams (Nagel and Rasmussen, 1994); price variations within stock-market trading (Bak et al., 1999); voting behaviors in elections (Kollman et al., 1992); identifying and exploring behavior in battlefields (Ilachinski, 1997); spatial patterns of unemployment (Topa, 2001); trade networks (Epstein and Axtell, 1996); business coalitions over industry standards (Axelrod, 2006); and the analysis of social networks of terrorist groups (North et al., 2004).
These examples can be construed as lying on a continuum, from minimalist academic models based upon ideal assumptions, to large scale commercial decision support systems based upon real-world data. Despite the advantages of ABM as a tool for simulation, it has not been widely adopted in geospatial research; although there is no obvious reason why this is the case.
Thomas Schelling is credited with developing the first social agent-based model in which agents represent people, and agent interactions represent a socially relevant process. Schelling (1971, 1978, see Figure 8‑4 and Figure 8‑12) demonstrated that stark segregation patterns can emerge from migratory movements among two culturally distinct, but relatively tolerant, types of households. Each cell represents a household, of ‘red’ or ‘black’ type in this example. Gray indicates that the cell is unoccupied at present. The basic rule for the evolution of the grid is that households ‘prefer’ to live in locations where the neighbors (8-fold or ‘Moore’ neighborhood) are of the same type. If a household is not ‘happy’ with its location, it may move elsewhere in the grid (to an unoccupied cell) or leave entirely. Even if the preference for having neighbors of the same type is quite modest (e.g. under 50%) Schelling found the process still resulted in strongly segregated groups, depending to some extent on the housing density. In the example illustrated the ratio of black to red households was set to 25% and the housing density was set to roughly 90% ― with lower housing densities evolution of a stable pattern occurs very rapidly. For discussion of a more sophisticated version of this kind of model, see Batty (2007) and Crooks, Castle and Batty (2007).
ABM did not begin to feature prominently in the geographical literature until the mid-1990s, when Epstein and Axtell (1996) extended the notion of modeling people to growing entire artificial societies. The goal was to understand the emergence of patterns, trends, or other characteristics observable in society or geography. Epstein and Axtell’s Sugarscape model demonstrated that agents could emerge with a variety of characteristics and behaviors suggestive of a rudimentary society (e.g. death, disease, trade, health, culture, conflict, war, etc.).
A. Initial state — random cell assignment
B. Stable evolved state
Epstein and Axtell describe Sugarscape as follows:
“Sugarscape simulates the behavior of artificial people (agents) located on a landscape of a generalized resource (sugar). Agents are born onto the Sugarscape with a vision, a metabolism, a speed, and other genetic attributes. Their movement is governed by a simple local rule: ‘look around as far as you can; find the spot with the most sugar; go there and eat the sugar.’ Every time an agent moves, it burns sugar at an amount equal to its metabolic rate. Agents die if and when they burn up all their sugar. Agents are accumulating sugar at all times, so there is always a distribution of wealth.”
The Schelling and Sugarscape models are important early examples of ABM ideas set within a rather theoretical framework. A typical, current generation ABM model, is illustrated in Figure 8‑5. This shows the a screenshot from a simulation of the movement of pedestrians (blue, red and green dots) within a subway (underground train station) hall area. This model, which may be run as a Java applet from the source site indicated or using the samples provided with the software distribution, incorporates:
|•||a realistic spatial framework|
|•||multiple passengers arriving and departing|
|•||multiple targets ― ticket machines, ticket booths, platforms, shops, exits …|
|•||free movement with obstacle avoidance|
|•||selectable simulation speed|
Models such as this provide valuable insights into spatial dynamics, in this case pedestrian behavior, which can clearly be applied in very many fields. They provide a medium where we can test ideas and hypothesis of phenomena which are not easy to do in the real world. One such example is pedestrian evacuation of a building during a fire. For example, without actually setting a building on fire we cannot easily identify people’s reactions to such an event. ABM, as with simulations in general, allow for such experiments. We can recreate the building in an artificial world, populate it with artificial people, who have a set of behaviors, start a simulated fire and watch what happens. Such simulations allow the modeler to identify potential problems such as bottle necks and allows for the testing of numerous scenarios.
Until recently the movement of agents has been dependent solely on the attributes of their immediate neighboring cells and their inhabitants, and their environments have not been based on real-world geographic features. Gimblett et al. (2002) were amongst the first to use real-world features. Their agent-based model was developed to evaluate the recreational use of Broken Arrow Canyon, Arizona. Specifically, current hiking, bike, and off-road trail paths were mapped in a GIS and potential alternatives simulated in order to aid management decisions of environmental protection and enhance recreational user experiences of the canyon. Dibble and Feldman (2004) have also extended the application of ABM to real-world environments by developing a three-dimensional extension to the Repast ABM toolkit. The extension has enabled the authors to model the control of infectious disease transmission, dynamics of civil violence, and coordination of social networks within three-dimensional landscape terrains, and social and spatial networks. More recently researchers have been exploring the potential of agent-based models in virtual worlds such as in Second Life (e.g. Crooks et al., 2008b) which provides opportunities for populating virtual worlds with spatial data and the creation of agent-based models which users can interact with.
This pioneering work has been extended to increasingly levels of real-world representation, with the development of very fine-scale models that merge ABM concepts with data from mainstream GIS and remote sensing, coupled with synthetic population creation (see further, below). Agents thus become idealized representatives of the actual population, located in a meaningful spatial and temporal framework. This approach has proved to be a particularly powerful tool in modeling disease spread in a range of environments - in the first example, illustrated in the video below (web version of this ebook only) the model explores the dynamics of refugee camps and the spread of diseases within such camps. The work carried out at George Mason University (GMU) involves modeling the Dadaab refugee camps which are located in Kenya, approximately 100 kilometers from the Somali border. The camps themselves are homes to roughly 500,000 people, with nearly 99% of the population coming from Somalia. Within the camps the mortality rate is ~ 0.44/10,000 per day with diseases such as cholera and measles being among the causes of death.
As noted above, recent technical advances in software and computing power have improved agent-based models in several respects, including:
|•||correct placement of all the main geographical features in a simulated area: buildings, roads, streams, parks etc. in a town or village - ensuring that the agents move within a realistic representation of the space under investigation. The use of ABM in conjunction with GIS software is now becoming more widespread. Readers who wish to explore this option using ESRI's ArcGIS software and Agent Analyst add-in are recommended to see the free resources (book, software, datasets) available at: http://resources.arcgis.com/en/help/agent-analyst/|
|•||dramatically improved model detail - the resolution of the modeled environment, the number of agents, the time intervals and the number of simulations, can now all be made very large, thereby making it possible to examine a very wide range of parameterizations and variations in model components such as initial conditions, positive and negative feedback and path dependency|
|•||inclusion of synthetic populations - because data is often only available at an aggregate level and agent based models operate at the individual level, there is great interest in generated synthetic populations (collections of individuals or entities) that match multiple attributes at various levels of aggregation. For example we may know the age, sex, gender, and many other variables at several levels of data zoning, but do not have information on the individuals that comprise these aggregate statistics. Such data can now be synthesized, using a variety of algorithms, such that the aggregate statistics for the set of simulated individuals matches those for each level of known aggregation. Indeed, in the US there are pre-built synthetic population now available for both human and some animal population. See: http://www.epimodels.org/ for more details. An example of the spatial allocations from part of the US human data synthetic set is shown below:|
As a second illustration of several of these concepts we take the example of an agent-based model created by Augustijn-Beckers et al. (2011). The team constructed an agent based model in order to simulate cholera diffusion in part of the city of Kumasi, in Ghana. They were particularly interested in comparing transmission mechanisms and constructed a model of part of the North-East section of the city that included:
|•||the spread of cholera from dump sites by the housefly (M. domestica) - vector transmission (VT)|
|•||runoff from dump sites as a pathway of bacteria and faeces to rivers - environment to human transmission (EH and HEH)|
|•||human to human transmission of cholera (HH), and|
|•||creation and use of a synthetic population of 3500 individuals representing appropriate age categories, income levels, blood groups (relevant for susceptibility to cholera) and other population dynamics like hygiene levels and access to pipe water|
In this model the geographic space was represented in considerable detail, with agents being individuals, synthetically generated and allocated into families that were in turn allocated to houses within the study area (447 in total). The figure below illustrates the kind of results achieved by the simulation.
Several useful observations have been made by the researchers regarding this particular study:
(i) by comparing simulated results with known outbreaks such models can provide guidance on steps that might be taken to reduce the risks of future outbreaks of cholera - e.g. by moving refuse dump sites, by increasing the number of properties served with piped water, by improved local drainage ditches etc.
(ii) increasing the extent of the study area would enable run-off modeling to be more realistic - the study area may not correspond to the relevant water catchment area
(iii) improvements to the hydrological model (e.g. river flow data - volume and velocity by time) would provide a more realistic representation of water flows, including surface runoff and potential well contamination; and
(iv) inclusion of non-resident agents (visitors) and dynamic behavioral changes (avoiding locations or households that have known infection) would add greater realism into the dynamics of the model