A neural network architecture composed of adaptively defined dissimilar single-neurons: applications in engineering design
- Τίτλος
- A neural network architecture composed of adaptively defined dissimilar single-neurons: applications in engineering design
- Θέμα
- Neural networks (Computer science)--Design and construction
- Δημιουργός
-
Neocleous, Constantinos C.
- Πηγή
- Higher Technical Institute
- Το πλήρες κείμενο είναι διαθέσιμο από το Υπουργείο Ενέργειας, Εμπορίου Βιομηχανίας και Τουρισμού.
- Εκδότης
- Library of Cyprus University of Technology
- Ημερομηνία
- 1997
- Δικαιώματα
- Απαγορεύεται η δημοσίευση ή αναπαραγωγή, ηλεκτρονική ή άλλη χωρίς τη γραπτή συγκατάθεση του δημιουργού και κάτοχου των πνευματικών δικαιωμάτων.
- Μορφή
- Γλώσσα
- en
- Τύπος
- text
- Αναγνωριστικό
- MED0598
- Σύνοψη
- The field of single-neuron models has been reviewed and systematically classified. A new generic single-neuron model structure is proposed that can serve as a parent model for most of the existing models. Three novel dynamical single-neuron models are presented and characterised for their response to different inputs. Based on these, new neural netWork architectures composed of dissimilar single-neuron models are proposed. The most general structure used is a feedforward in which the hidden layer is composed of two sublayers. These sub layers are made of different types of user selectable functional and dynamical single- neuron models. A set of MATLAB m-files has been developed for the building, training and recall of networks composed of dissimilar single-neuron models. The program allows for diverse network configurations and it enables a researcher to organise and conduct a rich set of simulations for the evaluation of the perfonnance of these architectures. Simulations on various benchmarks and on real engineering data for marine propellers and for neuromuscular disease classification have been conducted. The network showed a relatively faster rate of learning and better perfonnance compared to backpropagation based networks. It is however more complicated and hence slower in training. A model for the design of marine propellers has been prepared so that a naval architect may use it for propeller performance estimation and for future design appraisals.
- Πολυμέσα
- MED0598.pdf