Da Costa, L.; Landry, J.A. and Levasseur, Y. 2008. Treating Noisy Data Sets with Relaxed Genetic Programming. Lecture Notes in Computer Sciences. Springer-Verlag, Berlin, Heidelberg. Monmarché, N. et al. (Eds.): EA 2007, LNCS 4926, pp. 1–12.
Abstract
In earlier papers we presented a technique (“RelaxGP”) for improving the performance of the solutions generated by Genetic Pro- gramming (GP) applied to regression and approximation of symbolic functions. RelaxGP changes the definition of a perfect solution: in stan- dard symbolic regression, a perfect solution provides exact values for each point in the training set. RelaxGP allows a perfect solution to belong to a certain interval around the desired values. We applied RelaxGP to regression problems where the input data is noisy. This is indeed the case in several “real-world” problems, where the noise comes, for example, from the imperfection of sensors. We com- pare the performance of solutions generated by GP and by RelaxGP in the regression of 5 noisy sets. We show that RelaxGP with relaxation values of 10% to 100% of the gaussian noise found in the data can out- perform standard GP, both in terms of generalization error reached and in resources required to reach a given test error.