Abstract:
|
We wish to use large data sets to inform an individual-based model for learning about tree growth and forest dynamics over regional scales and more broadly. The computational complexity of our model and our regional scope make traditional approaches to Bayesian estimation and forecasting, e.g., MCMC, practically infeasible. To address the computational issue, we borrow perspective from the complex computer experiments literature to view our model as a complex simulator for which the goal is to create a computationally more tractable emulator model. From this perspective, our situation is relatively unique because we can run our simulator a relatively large number of times, and we have a large number of actual observations, two situations that are relatively rare in this context. Consequently, with many data and simulator runs, the familiar use of Gaussian processes as the basis of emulation becomes problematic. We explore alternative emulator formulations toward our goal to calibrate our simulator and predict tree growth.
|
ASA Meetings Department
732 North Washington Street, Alexandria, VA 22314
(703) 684-1221 • meetings@amstat.org
Copyright © American Statistical Association.