 |
    |
Controlling Overfitting in Symbolic Regression Based on a Bias/Variance Error Decomposition
Alexandros Agapitos, Anthony Brabazon, and Michael O’Neill
Financial Mathematics and Computation Research Cluster, Natural Computing Research and Applications Group, University College Dublin, Irelandalexandros.agapitos@ucd.ie
anthony.brabazon@ucd.ie
m.oneill@ucd.ie
Abstract. We consider the fundamental property of generalisation of data-driven models evolved by means of Genetic Programming (GP). The statistical treatment of decomposing the regression error into bias and variance terms provides insight into the generalisation capability of this modelling method. The error decomposition is used as a source of inspiration to design a fitness function that relaxes the sensitivity of an evolved model to a particular training dataset. Results on eight symbolic regression problems show that new method is capable on inducing better-generalising models than standard GP for most of the problems.
LNCS 7491, p. 438 ff.
lncs@springer.com
© Springer-Verlag Berlin Heidelberg 2012