Author: ANGEL MARIO GARCIA PEDRERO

Arquitectura neuronal apoyada en señales reconstruidas con Wavelets para predicción de series de tiempo caóticas

ANGEL MARIO GARCIA PEDRERO (2009)

Chaotic time series forecasting is a problem with a wide range of applications.

Although several methods have been developed to obtain an accurate prediction,

nowadays, forecasting is still an open problem. The chaotic nature of a time series

generator system makes difficult a prediction since it is highly sensitive to its initial

conditions; a small change may cause future estimations to diverges exponentially.

Traditional prediction techniques based on linear approximation methods fail

drastically when they are used to forecast this kind of time series. Some nonlinear

methods such as neural networks show a better performance in this cases, but an

accurate prediction is not possible yet, especially when they estimate many future

values.

In this work, a new architecture inspired on a recurrent neural network known

as Hybrid-Complex Neural Network for short time series forecasting is proposed

and its performance compared with another neural models. To achieve a prediction,

the proposed model takes advantage of information obtained by frequency

decomposition from wavelet transforms and of the oscillatory abilities of recurrent

neural networks. The proposed model shows a better performance than a fully recurrent

neural network and a feedforward neural network in forecasting 56 values

of the Sumsin (Misiti y colaboradores, 2008) time series and 11 of 111 series from

the NN5 dataset (Crone, 2008), the latter used as benchmark in several works and

in the “Artificial Neural Network & Computational Intelligence Forecasting Competition”.

The new model obtained an symmetric mean absolute percentage error

(sMAPE) average of 27.217%, compared to 60.755% obtained by fully recurrent

neural network and 49.276% obtained by a feedforward neuronal network.

La predicción de series de tiempo que presentan características caóticas es

un problema con un amplio rango de aplicaciones. Aunque se han desarrollado

gran cantidad de métodos para tratar de obtener una predicción precisa, éste

continúa siendo un problema abierto al día de hoy. La naturaleza caótica del

sistema generador de la serie dificulta la predicción al ser altamente sensible a las

condiciones iniciales; una pequeña perturbación puede provocar que los valores

subsecuentes diverjan exponencialmente.

Los métodos tradicionales de predicción basados en sistemas de aproximación

lineales fracasan fuertemente cuando se aplican para la predicción en este tipo de

series. Algunos métodos no lineales como las redes neuronales artificiales muestran

un mejor desempeño en estos casos; sin embargo, estos métodos aún no consiguen

una predicción suficientemente satisfactoria, en especial cuando estiman a la vez

varios valores del futuro.

En este trabajo se propone una nueva arquitectura inspirada en la red recurrente

conocida como red neuronal híbrida compleja (HCNN, del inglés Hybrid-

Complex Neural Network) para predicción a corto plazo de series de tiempo caóticas

y se analiza su desempeño comparándola con otros modelos neuronales. Para

lograr la predicción, el modelo propuesto aprovecha la información resultante de

la descomposición de frecuencias de la transformada wavelet y de las capacidades

oscilatorias de las redes neuronales recurrentes. El modelo propuesto mostró un

mejor desempeño que una red totalmente recurrente y una red de alimentación

progresiva, también conocida como red de alimentación hacia adelante (del inglés,

feedforward network), en la predicción de 56 valores de la series Sumsin (Misiti y

colaboradores, 2008) y 11 series de un conjunto de 111 series temporales llamado

NN5 (Crone, 2008) el cual es utilizado en varios trabajos de predicción de series

de tiempo a largo plazo y en la “Artificial Neural Network & Computational Intelligence

Forecasting Competition”. El nuevo modelo obtuvo un error porcentual

absoluto medio simétrico (sMAPE) promedio de 27.217% en la NN5, comparado

con 60.755% obtenido por la red neuronal totalmente recurrente y 49.276% por

la red de alimentación progresiva.

Master thesis

Measurement neural nets Discrete wavelet transforms Time series Time series forecasting CIENCIAS FÍSICO MATEMÁTICAS Y CIENCIAS DE LA TIERRA MATEMÁTICAS CIENCIA DE LOS ORDENADORES

Composite recurrent neural networks for long-term prediction of highly-dynamic time series supported by wavelet decomposition

MARIA DEL PILAR GOMEZ GIL ANGEL MARIO GARCIA PEDRERO JUAN MANUEL RAMIREZ CORTES (2010)

Even though it is known that chaotic time series cannot be accurately predicted, there is a need to forecast their behavior in may decision processes. Therefore several non-linear prediction strategies have been developed, many of them based on soft computing. In this chapter we present a new neural network architecutre, called Hybrid and based-on-Wavelet-Reconstructions Network (HWRN), which is able to perform recursive long-term prediction over highly dynamic and chaotic time series. HWRN is based on recurrent neural networks embedded in a two-layer neural structure, using as a learning aid, signals generated by wavelets coefficients obtained from the training time series. In the results reported here, HWRN was able to predict better than a feed-forward neural network and that a fully-connected, recurrent neural network with similar number of nodes. Using the benchmark known as NN5, which contains chaotic time series, HWRN obtained in average a SMAPE = 26% compared to a SMAPE = 61% obtained by a fully-connected recurrent neural network and a SMAPE = 49% obtained by a feed forward network.

Article

CIENCIAS FÍSICO MATEMÁTICAS Y CIENCIAS DE LA TIERRA FÍSICA ELECTRÓNICA ELECTRÓNICA

Neural networks and SVM-based classification of leukocytes using the morphological pattern spectrum

JUAN MANUEL RAMIREZ CORTES MARIA DEL PILAR GOMEZ GIL VICENTE ALARCON AQUINO JESUS ANTONIO GONZALEZ BERNAL ANGEL MARIO GARCIA PEDRERO (2010)

In this paper we present the morphological operator pecstrum, or pattern spectrum, as a feature extractor of discriminating characteristics in microscopic leukocytes images for classification purposes. Pecstrum provides an excellent quantitative analysis to model the morphological evolution of nuclei in blood white cells, or leukocytes. According to their maturity stage, leukocytes have been classified by medical experts in six categories, from myeloblast to polymorphonuclear corresponding to the youngest and oldest extremes, respectively. A feature vector based on the pattern spectrum, normalized area, and nucleus - cytoplasm area ratio, was tested using a multilayer perceptron neural network trained by backpropagation, and a Support Vector Machine algorithm. Results from Euclidean distance and k-nearest neighbor classifiers are also reported as reference for comparison purposes. A recognition rate of 87% was obtained in the best case, using 36 patterns for training and 18 for testing, with a three-fold validation scheme. Additional experiments exploring larger databases are currently in progress.

Article

CIENCIAS FÍSICO MATEMÁTICAS Y CIENCIAS DE LA TIERRA FÍSICA ELECTRÓNICA ELECTRÓNICA