A novel approach for knowledge extraction from Artificial Neural Networks

Abstract

Artificial Neural Networks (ANNs) have been used increasingly in recent years for numerous hydrological applications because of their ability to model the non-linear relationships. The major drawback of ANNs is that little is known about what is happening inside the ANN resulting into designating them as ‘black box’ models. Since long researchers have been trying to extract the knowledge from the trained ANNs, which will help them to become more widely accepted and reach their full potential as hydrological models. There were few such attempts in hydrology particularly for ANN model developed for river flow forecasting and rainfall–runoff process. The prime focus of this paper is to extract knowledge locked in the weights and biases of the trained ANN models using a new method proposed by the authors. For this, ANN models were developed to estimate pan evaporation at three stations in India using meteorological variables. The proposed method was able to throw a light on working of ANN and its understanding of physics in that it could correctly evaluate the influence of the input variables on the evaporation as directly or inversely proportional which was endorsed by the physics of the underlying process.

Keywords:

• Artificial Neural Network
• pan evaporation
• knowledge extraction