Accuracy tuning on combinatorial neural model

Abstract

The Combinatorial Neural Model (CNM) ([8] and [9]) is a hybrid architecture for intelligent systems that integrates symbolic and connectionist computational paradigms. This model has shown to be a good alternative to be used on data mining; in this sense some works have been presented in order to deal with scalability of the core algorithm to large databases ([2,1] and [10]). Another important issue is the prunning of the network, after the trainingp hase. In the original proposal this prunningi s done on the basis of accumulators values. However, this criterion does not give a precise notion of the classification accuracy that results after the prunning. In this paper we present an implementation of the CNM with a feature based on the wrapper method ([6] and [12]) to prune the network by usingt he accuracy level, instead of the value of accumulators as in the original approach.