@InProceedings{AnochiCampShigLuz:2014:DaAsAr,
author = "Anochi, Juliana Aparecida and Campos Velho, Haroldo Fraga de and
Shiguemori, Elcio Hideiti and Luz, Eduardo F. P. da",
affiliation = "{Instituto Nacional de Pesquisas Espaciais (INPE)} and {Instituto
Nacional de Pesquisas Espaciais (INPE)} and {Instituto de Estudos
Avan{\c{c}}ados (IEAv)}",
title = "Data assimilation with arti cial neural networks self-con guring
by MPCA",
booktitle = "Abstracts...",
year = "2014",
organization = "EngOpt.",
abstract = "Artificial Neural Networks (ANN) are computational techniques that
present a mathematical model inspired by the neural structure of
biological organisms, acquiring knowledge through experience,
which have been a technique successfully employed in many
applications on several research fields and currently under
intensive research worldwide. ANN with learning supervised have
emerged as excellent tools for deriving data oriented models, due
to their inherent characteristic of plasticity that permits the
adaptation of the learning task when data is provided. In addition
to plasticity, ANN also present generalization and fault tolerance
characteristics that are fundamental for systems that depend on
observational. Although much has been studied, there are still
many questions about the ANN models that need to be addressed. One
of the main issues of research in supervised ANN is to search for
an architecture optimum. In this paper, the determination of
optimal parameters for the neural network is formulated as an
optimization problem, solved with the use of meta-heuristic
Multiple Particle Collision Algorithm (MPCA). The MPCA
optimization algorithm emulates a collision process of multiple
particles greatly inspired on two physical behaviour inside of a
nuclear reactor absorption and scattering. The cost function has
two terms: a square difference between ANN output and the target
data (for two data set: learning process, and the generalization,
and a penalty term used to evaluate the complexity for the new
network architecture at each iteration. The concept of network
complexity is associated to the number of neurons and the number
of iterations in the training phase. In this work, two types of
neural networks are used, the radial basis function network (RBF)
and recurrent Elman. Here, the self-configuring networks are
applied to perform data assimilation to emulate the Kalman filter
is carried out with linear 1D wave equation.",
conference-location = "Lisbon",
conference-year = "2014",
label = "lattes: 2720072834057575 1 AnochiCampShigLuz:2014:DaAsAr",
language = "en",
targetfile = "Anochi_data.pdf",
url = "http://www.dem.ist.utl.pt/engopt2014/",
urlaccessdate = "12 maio 2024"
}