Session:

Evolutionary Computation and Neural Networks (ECNN)
Monday March 01, 2004, 12.00 – 12.20

Session Chair:

Paulo Cortez, Miguel Rocha

Paper Title:

Advanced Evolutionary Design of Generalized Recurrent Neural Networks

Author(s):

Dr. A.Dobnikar, University of Ljubljana, Slovenia
S. Vavpotic, University of Ljubljana, Slovenia

Abstract:

A new evolutionary algorithm for evolving generalized recurrent neural networks was developed. It has many advanced features, such as forking, exchanging mutation probability distributions and learning, which enable it to find optimal neural network topologies and weights for given problems. We also defined a new parameter, neural network processing speed, which enables us to use networks with one layer of neurons instead of those with many layers. It was proved that the new evolutionary algorithm always finds an optimal solution in a finite number of generations. The proposed algorithm was tested on different problem domains and the results obtained are very promising.