Application of multi-objective genetic algorithms to fitting piecewise linear models

Abstract

Despite the conflicting nature of low-complexity models versus error minimization in machine learning problems, the application of multi-objective learning algorithms is only recently acquiring an evident importance. In this article, an approach for piecewise linear regression is discussed. In particular, a multiobjective Genetic Algorithm is applied to creating a Pareto set of models, built by minimizing both the structural complexity of the models and the squared error of the output. Selection over this set of models is also discussed and one case example is presented that shows the performance of the algorithm. Moreover, a real case of daily temperature regression is studied. It can be concluded that the algorithm is capable of providing a near-optimal set of models that exhibit low regression errors and good generalization performance.