Competition for freshwater can be extreme - the future promises more of the same. Resource management agencies commonly seek to conserve aquatic biodiversity in ecosystems that provide freshwater for human use. Traditional modeling approaches and field experiments often cannot address real-world water allocation problems, particularly where managers seek to forecast outcomes under novel conditions. Mechanistic agent-based models can be well-suited for addressing these problems, but their application faces significant challenges. Key questions for the application of agent-based models include: What level of model complexity is appropriate? Can these complex models be ‘validated’ to foster their acceptance? We have addressed these questions via close linkage of model formulation and field studies, including multi-year monitoring efforts and focused field experiments.
Contrast of two stream fish models that included different levels of behavioral complexity corresponded similarly to multi-year observations of fish abundance and growth above and below a stream diversion. While this result suggests use of the simpler model - particularly when the difficulty of model calibration is considered - the behavioral flexibility included in the more complex model probably represents an important component of the response of the fish to environmental change. In a second exploration of the complexity issue, greater complexity in the form of multiple feeding modes was needed to match individual-growth results in a field experiment that included manipulation of streamflow.
Model ‘validation’ has been achieved in a narrow sense for the stream fish model by studies that allow model calibration parameters to be independently estimated through field observations. Values for parameters that describe baseline predation risk and food availability calibrated from multi-year abundance and growth have been supported by field observations. This correspondence has been valuable in gaining acceptance for the agent-based modeling approach where largescale, persistent manipulations that might validate model results at the population level are not available.