Usporedba unaprijednih neuronskih mreža s jednim i dva skrivena sloja u modeliranju linearnog dinamičkog sustava

undergraduate thesis

• Title (english): Comparison of feedforward neural networks with one and two hidden layers in modeling of linear dynamic system
• Author: Krešimir Milas
• Mentor: Dubravko Majetić (mentor)
• Granter: University of Zagreb; Faculty of Mechanical Engineering and Naval Architecture; Zagreb
• Defense date and country: 2015-09-21, Croatia
• Scientific / art field, discipline and subdiscipline: TECHNICAL SCIENCES; Mechanical Engineering

Abstract

U ovom se radu na primjeru učenja modeliranja linearnog dinamičkog sustava prvog reda (indentifikacija sustava) pokazala se kvaliteta učenja i generalizacije statičke neuronske mreže s jednim i dva skrivena sloja neurona. Za usporedbu se odabralo više različitih topologija jednoslojne i dvoslojne mreže. Uspoređivale su se mreže s jednakim brojem parametara učenja. Metode po kojima su uspoređivane su: brzina učenja do određenog NRMS-a te učenje do određenog koraka. Učenje se izvodilo po uzorku s zamahom prvog i drugog reda, a algoritam učenja je povratno raspostiranje greške. Nakon učenja gledalo se vrijeme potrebno da se dosegne cilj, postignut NRMS učenja te kvaliteta odziva. Nakon usporedba uzele su se mreže koje su se pokazale najbolje u generalizaciji te ih se pobudilo različitim pobudnim funkcijama (sinusnom i nagibnom) koje nisu bile u setu za učenje. Potom su ponovno uspoređene na temelju mogućnosti generalizacije odziva koje nisu učene.

Abstract (english)

This paper dealt with neural network learning of first lag order element. In it the quality of learning and quality of generalization of networks with one and two hidden layers is shown. Multiple topologies of networks with one and two layers were chosen for comparison. Networks with equal number of learning parameters were compared. There were two methods of evaluation of neural networks: learning up to a predetermined error threshold and learning up to predetermined number of epochs. Learning method used was pattern learning with momentum of first and second order. Upon completion of learning process time elapsed , achieved NRMS and quality of response were considered. After comparison, networks that proved best in generalization were introduced to different input functions (sine and ramp) which weren't in original learning set and compared again.

• Keywords: neuronske mreže; umjetna inteligencija; jednoslojne mreže; dvoslojne mreže; učenje po uzorku; zamah prvog reda; zamah drugog reda; generalizacija; P1 član; modeliranje sustava; povratno rasprostiranje greške; pobudne funkcije
• Keywords (english): neural networks; artificial intelligence; one hidden layer networks; two hidden layer networks; pattern learning; first order momentum; second order momentum; generalisation; first order lag element; system modelling; error backpropagation; input signal
• Language: croatian
• URN:NBN: urn:nbn:hr:235:176444
• Study programme: Title: Mechanical Engineering; Study programme type: university; Study level: undergraduate; Academic / professional title: sveučilišni/a prvostupnik/ prvostupnica (baccalaureus/baccalaurea) inženjer/inženjerka strojarstva (sveučilišni/a prvostupnik/ prvostupnica (baccalaureus/baccalaurea) inženjer/inženjerka strojarstva)
• Type of resource: Text
• File origin: Born digital
• Access conditions: Open access
• Created on: 2019-05-22 14:37:36