Artificial neural networks employing fuzzy numbers for rapid risk estimation

This paper aims in the construction of Artificial Neural Networks that estimate pure real numbers in the closed interval [0,1] representing the overall degree of risk in a multi attribute decision making case. These numbers that represent Fuzzy Sets, take into consideration all of the factors that have uneven influence in the determination of the risk degree. Of course this could have been achieved by a single ANN, but this research aims in viewing the problem under different perspectives performing an optimization analysis. In this way it can be determined if a case characterized by average values for all the parameters is more risky than another one that has extremely high values for some parameters (no matter the values of the others).This modeling effort establishes a direct cooperation between an Intelligent Fuzzy Decision Support System with Artificial Neural Networks by using the output of the Fuzzy System to train and test the Neural Network. This means that both the input and the output of the Neural Networks are Fuzzy numbers. The ANN is very useful when only rough estimations of the involved attributes are available and a rapid risk evaluation is required. The produced model has been applied for the determination of the Overall Degree of Torrential Risk in the mountainous watersheds of the Greek island “Samothraki”. The paper suggests the clustering of cases under risk evaluation and the construction of a prototype risk estimation neural model for each cluster.