مقايسه‌ي روش‌هاي مختلف ارزيابي ميزان اثرگذاري متغيرهاي ورودي در شبكه‌هاي عصبي پيشخور با استفاده از داده‌هاي ژيوتكنيكي‌

A COMPARISON BETWEEN DIFFERENT METHODS OF ASSESSING THE IMPACT OF INPUT FEATURES IN FEEDFORWARD NEURAL NETWORKS USING GEOTECHNICAL DATABASE

The ba ck-propagation neural network (BPNN) has been researched and applied as a convenient t ool in a variety of application areas in civil engineering. In particular, BPNN has been applied to many geotechnical engineerin g problems and has demonstrated some degree of success. A review of the lit erature reveals th at BPNN has been used successfully in pile capacity prediction, modeling soil behavior and liquefaction , etc. However , learning algorithms, such as the BPNN, do not give information on th e effect of each input parameter or influencing variable upon t he predicted output variable. In other words , it is not possible to find out immediat ely how the weights of the network or the activation values of the hidden neurons are related to the set of data being handled. Instead, ANNs have been pre sented to the user as a kind of ʹblack box ʹ, whose ext remely complex work transform s inputs into predetermined output s. To deal with this problem, during th e last 10 years, differ ent interpretative methods for analyzing the effect or importance of input variables on the output of a feedforward neural network have been proposed . In th is paper, six methods that give the relative cont ribution of the input factors were compare d. The data used for training the networks is based on the laboratory test s for determining th e dynamic properties of aggregate-clay mixtures. Finally, the method which best interpret s the networ ks is introduced .