L. Ferariu - Academia.edu (original) (raw)
Uploads
Papers by L. Ferariu
The Bulletin of Politechnical Institute of Iasi, Jan 1, 2010
This paper presents a new genetic programming based approach for the design of feed forward hybri... more This paper presents a new genetic programming based approach for the design of feed forward hybrid neural models. A special encryption of the neural model as a directed acyclic graph (DAG) is suggested, intended to exploit the modularity of the neural topology. This allows for a flexible development of partially interconnected neural structures, having heterogeneous layers with both Gaussian and perceptron neurons. Customized compatible genetic operators, aimed to work simultaneously on the neural architecture and parameters, provide an efficient exploration of the search space. Additionally, the algorithm makes use of a back-propagation procedure, employed as a local Lamarckian optimization, for a faster computation of model parameters. The performances of the suggested approach are illustrated on the identification of an industrial subsystem from the Sugar factory of Lublin, Poland.
The Bulletin of Politechnical Institute of Iasi, Jan 1, 2010
This paper presents a new genetic programming based approach for the design of feed forward hybri... more This paper presents a new genetic programming based approach for the design of feed forward hybrid neural models. A special encryption of the neural model as a directed acyclic graph (DAG) is suggested, intended to exploit the modularity of the neural topology. This allows for a flexible development of partially interconnected neural structures, having heterogeneous layers with both Gaussian and perceptron neurons. Customized compatible genetic operators, aimed to work simultaneously on the neural architecture and parameters, provide an efficient exploration of the search space. Additionally, the algorithm makes use of a back-propagation procedure, employed as a local Lamarckian optimization, for a faster computation of model parameters. The performances of the suggested approach are illustrated on the identification of an industrial subsystem from the Sugar factory of Lublin, Poland.