Konsep statistika sebagai kriteria pemberhentian pelatihan jaringan saraf tiruan untuk mengatasi keterbatasan data (studi kasus pada prediksi tegangan permukaan Surfaktan MESA dari minyak kelapa sawit)

Abstract

There exist situations such as marine ecology where artificial neural network (ANN) implementation must cope with scarce data. Efforts to add more data in this sort of field will most likely be overlooked for being expensive and time consuming. Cross-validation comes across the opportunity of still using ANN to survive this poor data condition while not compromising its generalization. This paper proposed prediction output interval as stopping-criteria employed in 5-fold validation and demonstrated its performance when selecting best hidden layer number of neurons to be used for surfactant-MESA surface tension prediction based on only ten data pairs. Output interval building, derived from statistics concept, was gaining from the fact that there were two experiment repetitions from the case which could be generalized to cases with two measurement repetitions. Repetitions are expected to produce difference. A 95% confidence interval to estimate mean from these difference samples was then used to create output interval for ANN training. Backpropagation ANN with this added stopping criterion successfully gave 1.83 x 10-3 cross-validation MSE comparable to 2.72 x 10-3 of overfitted result, using only epoch and gradient of training MSE as stopping criteria. A slight variation on the width of the output interval, using sum of sample mean and standard deviation, also gave 3.04 x 10-4 MSE as best results. These intervals were found to perform well when set as constants, drawn from parametric statistical figures, around two times standard deviation of surface tension measurement differences from the experiment repetitions.