@InProceedings{CortivoChalCamp:2012:CoMLAd,
author = "Cortivo, Fabio Dall and Chalhoub, Ezzat Selim and Campos Velho,
Haroldo Fraga de",
affiliation = "{Instituto Nacional de Pesquisas Espaciais (INPE)} and {Instituto
Nacional de Pesquisas Espaciais (INPE)} and {Instituto Nacional de
Pesquisas Espaciais (INPE)}",
title = "A committee of MLP with adaptive slope parameter trained by the
quasi-Newton method to solve problems in hydrologic optics",
booktitle = "Proceedings...",
year = "2012",
pages = "1--8",
organization = "International Joint Conference on Neural Networks, (IJCNN).",
publisher = "Institute of Electrical and Electronics Engineers",
address = "Piscataway",
note = "{Setores de Atividade: Educa{\c{c}}{\~a}o.}",
keywords = "Hydrologic optics, Multi layer perceptron, Phase functions,
Quasi-Newton methods, Single scattering albedo, Artificial Neural
Networks, Multilayer Perceptron, Backpropagation, Quasi-Newton
Method, hydrologic optics, Single Scattering Albedo.",
abstract = "Artificial Neural Networks (ANNs) can be used to solve problems in
Hydrologic Optics. A relevant problem is the estimation of the
single scattering albedo and the phase function parameters, from
the emitted radiation at the surface of natural waters. In this
work we use a committee of ANNs of Multilayer Perceptron type to
perform the estimation of the two mentioned parameters. The
training of each network is formulated as a nonlinear optimization
problem subject to constraints. In addition, each activation
function has a distinct slope parameter, that is initially chosen
by a random number generator function. This set of parameter
(slopes) was included within the free variables network set in
order to be adjusted to reach optimal values, together with the
weights and biases, during the network training. This procedure
(slope parameters inclusion) makes each one of the activation
functions to have a different slope. Each network that composes
the committee was trained independently, in order to become expert
for the estimation of only one of the hydrologic parameters. For
the networks training, we used the quasi-Newton method that is
implemented in E04UCF subroutine, in the NAG library, developed by
the Numerical Algorithms Group - NAG. The use of the quasi-Newton
method to train the networks together with the distinct slope
parameters resulted in a network with a fast learning and
excellent generalization. Once the networks were trained, they
were grouped so to share the input patterns, but remained
independent from one another. For the validation/generalization
test we used two distinct sets. For all considered noise levels,
we obtained 100% of correct answers for the first set, and above
90% of correct answers for the second se.",
conference-location = "Brisbane",
conference-year = "10-15 June 2012",
doi = "10.1109/IJCNN.2012.6252665",
url = "http://dx.doi.org/10.1109/IJCNN.2012.6252665",
isbn = "978-146731490-9",
issn = "1098-7576",
label = "lattes: 8068157900374950 2 CortivoChalVelh:2012:CoMLAd",
language = "en",
organisation = "IEEE Computational Intelligence Society (CIS); International
Neural Network Society (INNS",
targetfile = "cortivo_committee.pdf",
urlaccessdate = "30 abr. 2024"
}