Selection pressure driven sliding window behavior in genetic programming based structure identification

Virtual sensors are a key element in many modern control and diagnosis systems, and their importance is continuously increasing; if there are no appropriate models available, virtual sensor design has to be based on data. Structure identification using Genetic Programming is a method whose ability to produce models of high quality has been shown in many theoretical contributions as well as empirical test reports. One of its most prominent shortcomings is relatively high runtime consumption; additionally, one often has to deal with problems such as overfitting and the selection of optimal models out of a pool of potential models that are able to reproduce the given training data. In this article we present a sliding window approach that is applicable for Genetic Programming based structure identification; the selection pressure, a value measuring how hard it is to produce better models on the basis of the current population, is used for triggering the sliding window behavior. Furthermore, we demonstrate how this mechanism is able to reduce runtime consumption as well as to help finding even better models with respect to test data not considered by the training algorithm.