engleski

Synthetic method back propagation analytic hierarchy process

The article shows limitatio s of the method AHP which are the co seque ces of li ear depe da ce of the output variable upo i put variables. For this purpose we compare two ways of modelli g profit fu ctio : usi g eural etwork a d usi g AHP. Each of the quoted methods has some adva tages: eural etworks have a possibility of lear i g from data a d modelli g o li earity; o the other ha d, i AHP method parameters are modelled by people - experts. The data base of cases solvi g the problem of supplyi g credits i ba ks is used to prove the limitatio s of AHP method. It is proved that AHP is the special case of Neural Network Back - Propagatio with hierarhic levels where the ide tity is the tra sfer fu ctio . Further, it is showed that it is possible to reduce the hierarhical AHP etwork to the eural etwork with o ly o e hidde level a d the ide tity as the tra sfer fu ctio . The exact ess of modelli g the classical BP eural etwork is compared with the li ear AHP etwork. Furthermore, it is showed that for the typical case of supplyi g credits AHP etwork is less exact tha BP etwork. The i exact ess of BP etwork is the result of oise i lear i g data a d the imprecisio of modelli g the profit fu ctio ; o the other ha d, the i exact ess of AHP is the result of oise i data that experts give a d o li ear depe da ce of the profit fu ctio upo data. To u ite good properties of AHP a d BP method the example is give i which the Neural Network BP lear s through the data that we get as the result of applyi g the AHP method.

back propagation; AHP; analytic hierarchy process

nije evidentirano