Authors:
M. P. Cuéllar
;
M. Delgado
and
M. C. Pegalajar
Affiliation:
University of Granada, Spain
Keyword(s):
Recurrent Neural Network, hybrid algorithms, time series.
Related
Ontology
Subjects/Areas/Topics:
Artificial Intelligence
;
Artificial Intelligence and Decision Support Systems
;
Biomedical Engineering
;
Biomedical Signal Processing
;
Computational Intelligence
;
Enterprise Information Systems
;
Evolutionary Programming
;
Health Engineering and Technology Applications
;
Human-Computer Interaction
;
Methodologies and Methods
;
Neural Network Software and Applications
;
Neural Networks
;
Neurocomputing
;
Neurotechnology, Electronics and Informatics
;
Pattern Recognition
;
Physiological Computing Systems
;
Sensor Networks
;
Signal Processing
;
Soft Computing
;
Theory and Methods
Abstract:
Dynamical recurrent neural networks are models suitable to solve problems where the input and output data may have dependencies in time, like grammatical inference or time series prediction. However, traditional training algorithms for these networks sometimes provide unsuitable results because of the vanishing gradient problems. This work focuses on hybrid proposals of training algorithms for this type of neural networks. The methods studied are based on the combination of heuristic procedures with gradient-based algorithms. In the experimental section, we show the advantages and disadvantages that we may find when using these training techniques in time series prediction problems, and provide a general discussion about the problems and cases of different hybridations based on genetic evolutionary algorithms.