Abstract:
|
In recent years, wildfires have devastated communities and represent an immediate danger to humans in terms of property damage and death. There is evidence that wildfires are becoming more frequent and are increasing in size. Moreover, wildfires are predicted to continue to increase in a warming climate. Prediction of wildfires can be difficult when one acknowledges that firebrand showers are known to be a complex nonlinear stochastic process. Motivated by an extension of the Kolmogorov Arnold representation theorem, we develop a new class of Bayesian neural network (BNN) models to analyze unknown nonlinear functions that treat the activation functions as unknown. Traditional BNNs have considerably more computational difficulties than standard neural network models that make use of efficient backpropagation algorithms. We impose conditional independence assumptions among the activation functions that lead to an efficient implementation of our BNN. This particular type of BNN can be interpreted as a nested or "deep" hierarchical generalized transformation model (DHGT). An analysis of a recent California wildfire using a DHGT is presented.
|