12th INTERNATIONAL STATISTICS DAYS CONFERENCE, İzmir, Türkiye, 13 - 16 Ekim 2022, ss.42
Multilayer feed forward neural networks have been widely used for prediction, forecasting
and classification over the past few years. However, it is a known fact that the mostly
preferred Mc - Culloch Pitts neuron model used in these network types does not give a
successful prediction performance in data sets with outliers. Therefore, robust neuron models
using median and trimmed mean aggregation functions have been proposed. However, these
studies were generally focused on time series forecasting. In this study, we developed new
neuron models using NO estimator and the Winsorized mean for prediction, classification,
and time series forecasting. NO is a quantile estimator with weights determined by using a
subsampling approach. For estimating a population quantile, it uses all order statistics in a
sample and the accompanying weights of the order statistics are calculated from a Binomial
Distribution. The proposed NO and Winsorized mean neuron models are not sensitive to
outlying observations. Back propagation, particle swarm optimization and artificial bee
colony optimization algorithms were used when training multilayer neural networks and
several activation functions such as sigmoid, hyperbolic, tangent, and rectified linear unit
were tried. All steps of the study were performed using statistical programming language R.
The written functions of the proposed neural networks enable the prediction of new
observations and observing the change of errors at each iteration by providing a dynamic plot.
The developed methods were applied on the real data sets and their performances were
compared with the existing ones. More successful results were achieved in terms of different
performance criterions