An Evolutionary Algorithm for Inversion of ANNsHenrik Jacobsson and Bj orn OlssonDept

Before using a trained artiicial neural network (ANN) in an application it is important to identify inputs which cause incorrect behaviours. We therefore propose the use of an evolutionary algorithm (EA) to invert the mappings of ANNs. The EA is used to search for input patterns which produce strong (distinct) classiications into one of the classes. Since the input space is typically very large, multimodal, and poorly understood, EAs are likely to be more robust than gradient methods , with a lower probability of getting stuck on local optima. Analysis of our results supports this hypothesis. Our evolutionary algorithm also involves the use of niching, which allows it to simultaneously explore multiple regions of the search space. The resulting population of input patterns therefore typically represents a set of distinctly diierent instances. This property is important, since the aim is to identify inputs which are erroneously classiied. We show how analysis of the set of inputs found by inversion can lead to detection of aws in the ANN, and we discuss the possibilities of using this inversion method as a tool for validation and retraining .