Pruned Committee Neural Network Based on Accuracy and Diversity Trade-off for Permeability Prediction

Abstract

Committee Machine (CM) or ensemble introduces a machine learning technique that aggregates some learners or experts to improve generalization performance compared to single member. The constructed CMs are sometimes unnecessarily large and have some drawbacks such as using extra memories, computational overhead, and occasional decrease in effectiveness. Pruning some members of this committee while preserving a high diversity among the individual experts is an efficient technique to increase the predictive performance. The diversity between committee members is a very important measurement parameter which is not necessarily independent of their accuracy and essentially there is a tradeoff between them. In this paper, first we constructed a committee neural network with different learning algorithms and then proposed an expert pruning method based on diversity and accuracy tradeoff to improve the committee machine framework. Finally we applied this proposed structure to predict permeability values from well log data with the aid of available core data. The results show that our method gives the lowest error and highest correlation coefficient compared to the best expert and the initial committee machine and also produces significant information on the reliability of the permeability predictions.