Modified Back Propagation Algorithm for Learning Artificial Neural Networks

Abstract

The Bacl:propa.g,auon algorithm has plaved an 1mportan1 part in lnlining mu1ttlay�

Artificial Neural Networl:s. It has II mathem.11,e.11 found.a!ioo th.at i,, su-ong ,f not

lughly peectical Dc,pite ilS lmiita1ion, Badcpropagatiou has dnmatically expanded

• the range of problems lo which Artificial Neural N�rks Cll1 be appt,cd

Bui there exist many chsadVlllluagcs and limitatioru; in the bacl<pn,paganon

algonthm such as

I. The optimal vduc of learning rate is 00( dr:lcrm.i11cd e,ualy by marbemancal

rule, L>ur. uc determlncd by trials

2. The inJu,l values ofweighls ind biases are not dewmintd c�1cdy by matl>cmatical rule., but are dele1mincd by triab 3. Thoe optim1l rltJmhet ofhidden layer is not dnerrntne.:I -tly

4. Necc:1s.ry number of hidden -roos is •IUI Jt1ffl!Un«I e:<.aal)'

This thesis tne5 to sol,e !Oll\e of the above problemi :tm "1111.:.es od,tt modification ,n the 11teepe:st descent method, which is the mam ide.:, of1he bDclrp�gation algori1hm • that &hou!d make 1h,e b«kprop1gauon algon1hm more @ellef3I and try to elimin11e some heuristics eici�111g in the algonttun.