Bayesian Inference of Computationally Expensive Fisheries Models

Complex fisheries models, such as individual-based or spatially-explicit models, provide a potential framework for exploring complex dynamics in populations, communities and ecosystems.  Inference that ignores individual variability and/or spatial complexity may provide biased, imprecise or overly-precise platforms for management advice.  However, standard inference is often not possible in such computationally expensive models.  Bayesian emulators are a method with the capacity to make inference about such models.  An emulator is a stochastic representation of a simulator conditioned by evaluations of that simulator at known inputs.  The emulator allows us to interpolate (or extrapolate) the evaluations of the simulator to beliefs about the simulator output for any input and conversely to make inferences about the “best inputs”, conditional on a given data set.  We have extended and adapted this method so that it may be used to make inference about computationally expensive fisheries problems.  While the application of this method and its potential for inference of such models is exciting, there are other potential uses for the Bayesian emulation framework in fisheries management.  For example, conditioned emulators could be used to provide more accurate inference about fish stocks between stock assessment years rather than relying on projections alone.