The popularity of agent-based modeling has exploded in the past decade – and for good reason. More and more researchers are recognizing that human decision-making – as messy and unpredictable as it seems – is an incredibly powerful driver of human and natural system dynamics. Agent-based models (ABMs) offer the means to test our understanding of decision-making processes and their consequences with more sophisticated (and often more realistic) simulations. Despite this growing popularity, though, ABMs introduce new and profound challenges in their use, testing, and interpretation.
A question that I am confronted with regularly – and have been thinking a lot about recently – is this: “What are the spatial and temporal scales of human-decision-making relevant to the land change phenomenon I want to study?”
I will likely return to this question in some form or another many times on this blog, but lately I am concerned with the data dimensions of this question. For example, if one wants to model household-driven land use changes in frontier regions (e.g. Parker et al., 2008), then land-use decisions at the household level are probably your likely target. However, what if you want to apply this model over a large region? With lots of agents? Or have it be general enough to apply across many sites in different land-use systems ( a la virtual laboratory!)? Then the relevant level of decision-making becomes murkier, especially when considering the data that will be needed to parameterize and test the model. In this example, you might need to combine household surveys, regional land-use/cover data, census data, and perhaps some global data such as market influence (Verburg et al., 2006).
Much has been written on this topic, and as there are many facets to this challenge, there are also many different ways to address it. The various perspectives on this issue seem to consistently fall into categories:
- How to coordinate different sources of data?
- What types of data should be used to test the model?
- What is the focal scale of analysis?
These questions are often raised in the context of using ABMs in combination with a geographic information system (GIS). Crooks and colleagues (2008), Parker (2005), and Brown and colleagues (2005) are good sources to begin answering these questions.
For our purposes here, I’ll ask a simpler question: How does one get these different kinds of data that cover different scales and resolutions to play nicely together?
Coordinating the scale and/or resolution of spatial data in a GIS with those of the decision-making processes in an ABM can be quite tricky. So, as the modeler, you have to make some choices about how closely the model should be “coupled” with the GIS: 1) Spatial data layers can be used simply to initialize an ABM by parameterizing the simulated landscape; 2) GIS and the ABM can be “loosely coupled” such that spatial data files are updated independently and then passed back and forth at designated times; or 3) the ABM can be fully embedded within the GIS (e.g. the new Agent Analyst tool with Repast). The choice of either of these methods will depend on factors such as the research question and the computational expense of passing files between systems. Whichever path one chooses, though, the challenge of coordinating low resolution spatial data with fine-grain social data remains.
Arguably a more important consideration, then, is the level of abstraction within the model. I have recently toyed with this question in a model in which my agents represent aggregations of households in settlements rather than individual households. This, of course, has some conceptual limitations. However, it has important conceptual and practical advantages as well; in particular, the ability to represent a landscape-level entity generally by avoiding some social interactions among households that are notoriously difficult to generalize. This representation is certainly controversial (and probably nonsense to some household livelihoods researchers), but considering the types of data needed to address regional land-use questions, an agent representation that works nicely with global datasets is not entirely a bad thing.
Thus, I will plant this seed: In what situations does a settlement-level agent act as a reasonable representation of land-use decision-making? This is an ongoing topic of research for me, and will undoubtedly lead to subsequent posts. Stay tuned!
Brown, D.G., Riolo, R., Robinson, D.T., North, M., and Rand, W. (2005). Spatial process and data models: Toward integration of agent-based models and GIS. Journal of Geographical Systems, 7:25–47.
Crooks, A. T., Castle, C. J. E., and Batty, M. (2008), Key Challenges in Agent-Based Modelling for Geo-spatial Simulation. Computers, Environment and Urban Systems, 32(6): 417-430.
Verburg PH, Ellis EC, Letourneau A (2011) A global assessment of market accessibility and market influence for global environmental change studies. Environ Res Lett, 6.