Adaptive neural network ensemble using prediction frequency

Abstract Neural network (NN) ensembles can reduce large prediction variance of NN and improve accuracy. For highly non-linear problems with insufficient data set, the accuracy models becomes unstable, resulting in a decrease ensembles. Therefore, this study proposes frequency-based ensemble that identifies core values, which are members to be used expected concentrated near true response. The classifies values ​​supported by multiple ​​by conducting statistical analysis frequency distribution, is collection ​​obtained from various for given point. searches range contains above certain frequency, thus predictive performance improved excluding low ​​and coping uncertainty most frequent value. An adaptive sampling strategy sequentially adds samples based on calculated as proposed efficiently. Results case studies show higher than Kriging other existing methods. In addition, effectively improves compared previously developed space-filling variance-based strategies.