Improved Time Series Prediction and Symbolic Regression with Affine Arithmetic
Venue
Genetic Programming Theory and Practice IX, Springer, 233 Spring Street, New York, NY 10013 (2011), pp. 97-112
Publication Year
2011
Authors
Cassio Pennachin, Moshe Looks, J. A. de Vasconcelos
BibTeX
Abstract
We show how affine arithmetic can be used to improve both the performance and the
robustness of genetic programming for problems such as symbolic regression and time
series prediction. Affine arithmetic is used to estimate conservative bounds on the
output range of expressions during evolution, which allows us to discard trees with
potentially infinite bounds, as well as those whose output range lies outside the
desired range implied by the training dataset. Benchmark experiments are performed
on 15 symbolic regression problems as well as 2 well-known time series problems.
Comparison with a baseline genetic programming system shows a reduced number of
ļ¬tness evaluations during t raining and improved generalization on test data,
completely eliminating extreme errors. We also apply this technique to the problem
of forecasting wind speed on a real world dataset, and the use of affine arithmetic
compares favorably with baseline genetic programming, feedforward neural networks
and support vector machines.
