Nonparametric Radial Basis
Function Regression Using
Genetic Algorithms and Model Averaging
|
Andrea
Nibbe, Statistics Department, Colorado State University
Thursday, 11 December 2003
3:10 PM
E106 Engineering Building
ABSTRACT
This work
proposes a model selection approach to radial basis functions where the
number and locations of basis functions is treated as an unknown. The
genetic algorithm is used to optimize the basis function placements, where the measure of a
model's performance is defined by Akaike's Information criterion (AIC). Predictions are made by applying model averaging
strategies to the top performing models sampled throughout the genetic
algorithm search.
The
most promising of the model averaging strategies involves fitting a
weighted least squares regression to the predicted values from the top
models. The input weights are
the inverse of the AIC value from the corresponding model fit. Simulation
results show that the proposed method outperforms a Bayesian approach to
the same problem.
|
