Geocomputation

Developing and Testing Simulation Models

Geographic simulation models may be cellular automata tessellations of continuous landscapes to [Page 168]simulate processes such as erosion or wildfire; agent-based models on continuous landscapes to simulate recreational behavior, such as hiking, or wildlife behavior, such as flocking or grazing; or agent-based models on networks to simulate processes, such as human migration or travel among cities. Hybrid models combine aspects of each, such as graphical cellular automata, where nodes on a graph update their states according to cellular-automata rules regarding the states of neighboring nodes, or urban models, where sophisticated cellular automata represent changes in land use and land cover, while heterogeneous mobile agents represent business or residential location choices of urban residents. Alternatively, systems of mathematical equations can be appropriate for modeling specific behavioral rules or for representing complementary processes that have predictable responses, such as vegetation growth, evaporation, or similar biogeophysical transformations.

Simulation models can be classified as deterministic or stochastic. Deterministic simulation models always generate the same output for any given set of rules and initial conditions. Stochastic simulation models allow for effects of chance events, which may accumulate and interact to generate myriad simulation output results for any given set of rules and initial conditions. Random number seeds control one or more random number series to simulate one or more different types of chance events in stochastic models.

All simulation types share common principles regarding rigorous design, development, verification, calibration, and validation. Except for the most trivial examples, models of all kinds are necessarily simplifications of the complex distributed systems they represent. Although it may seem counterintuitive, models are usually most useful when they are designed to be the simplest possible representations capable of generating the phenomena we seek to study. While more complicated models may appear better, this is generally due to overfitting only to a particular data set. In contrast, simpler models can provide valuable insights regarding the behavior of similar systems elsewhere or in the future.

...