A Stratified Sampling Algorithm for Artificial Neural Networks

Abstract

Artificial Neural Networks (ANN) MultiLayer Perceptron (MLP) are widely applied in a variety market segments to handle with real complex problems. The ability to deal with tasks in real time is essential in an environment that uses large volume do information available. In each new project, a decision-making system using ANN with time reduction and data processing is a key issue to test various learning algorithms; containing a variety of parameters when using this technology. From this starting point, the MLPs used data collected from a specific phenomenon and, based on statistical estimators, applied a data extraction algorithm for stratified sampling, aiming to reduce the time of ANN processing. In this context, this work proposes a Stratified Sampling algorithm (SSA), which was developed to minimize processing MLPs time without losing coverage and assertiveness, when comparing with training conducted on a population database. The case study consisted of a ANN performance influence with a population database and with its sample data obtained by the SSA model. This procedure with the RNAs aimed to evaluate the following properties: (1) meet the pre-established criteria of reliability of the model; (2) have a computer-automated procedure; (3) sort and select records more correlated, and (4) maintain sampling results within a track of assertiveness of total results obtained. From the realization of this case study, it was possible to identify the following gains made by the (1) reduction of ANN processing time by providing: (2) optimization of processing time; (3) automatic network selection; and (4) automatic parameters selection for training algorithms.