pdf139.pdf

Embed Size (px)

Citation preview

  • 8/11/2019 pdf139.pdf

    1/6

    139 1

    AbstractIn this work an algorithm to adjust parameters

    using a Bayesian method for cumulative rainfall time series

    forecasting implemented by an ANN-filter is presented. The

    criterion of adjustment comprises to generate a posterior

    probability distribution of time series values from forecasted time

    series, where the structure is changed by considering a Bayesian

    inference. These are approximated by the ANN based predictor in

    which a new input is taken in order for changing the structure

    and parameters of the filter. The proposed technique is based on

    the prior distribution assumptions. Predictions are obtained by

    weighting up all possible models and parameter values accordingto their posterior distribution. Furthermore, if the time series is

    smooth or rough, the fitting algorithm can be changed to suit, in

    function of the long or short term stochastic dependence of the

    time series, an on-line heuristic law to set the training process,

    modify the NN topology, change the number of patterns and

    iterations in addition to the Bayesian inference in accordance with

    Hurst parameter H taking into account that the series forecasted

    has the same H as the real time series.

    The performance of the approach is tested over a time series

    obtained from samples of the Mackey-Glass delay differential

    equations and cumulative rainfall time series from some

    geographical points of Cordoba, Argentina.

    I ndex Terms Bayesian approach, Neural networks, timeseries forecast, Hursts parameter.

    I. INTRODUCTION

    he ANNs are mostly used as predictor filter with an

    unknown number of parameters performed by a lot of

    author, recently, such as in [2][13][16]. In turn, in this

    Manuscript received April, 18, 2012. This work was supported in part by

    Universidad Nacional de Cordoba, FONCYT-PDFT PRH N3 (UNC Program

    RRHH03), SECYT UNC, National University of Catamarca, National Agency

    for Scientific and Technological Promotion (ANPCyT), Departments ofElectrotechnics FCEFyN of Universidad Nacional de Cordoba.

    C. Rodriguez Rivero is with Universidad Nacional de Cordoba, Department

    of electrical and electronic engineering, Faculty of Exacts, Physical and Natural

    Sciences, Cordoba, Argentina (phone 0054-351-4334147; e-mail:

    [email protected]).

    J. A. Pucheta is with Universidad Nacional de Cordoba, Department of

    electrical and electronic engineering, Faculty of Exacts, Physical and Natural

    Sciences, Cordoba, Argentina (phone 0054-351-4334147; e-mail:

    [email protected]).

    M. R. Herrera is with National University of Catamarca, Department of

    electronic engineering, Faculty of Technology and Applied Sciences, (e-mail:

    [email protected]).

    work the main purpose is to estimated water availability

    useful for control problems in agricultural activities such

    as seedling growth and decision-making. The number of

    filter parameters is set in function of the roughness of the

    time series. These are considered as random variables

    whose distribution is inferred by posterior probability

    from the data, in which is included as an additional

    parameter, the number of hidden neurons and modeling

    uncertainty[1]. The model selectionproblem is, therefore,unavoidable; researchers must decide which model is best

    at summarizing the data for each task of interest.The Bayesian approach permits propagation of uncertainty

    in quantities which are unknown to other assumptions in the

    model, which may be more generally valid or easier to guess in

    the problem. For neural networks, the Bayesian approach was

    pioneered in [2]-[3], and reviewed [5], [6]and [7]. The main

    difficulty in model building is controlling the complexity of the

    model. It is well known that the optimal number of degrees of

    freedom in the model depends on the number of training

    samples, amount of noise in the samples and the complexity of

    the underlying function being estimated.

    The procedure of determining the prior density and

    likelihood functions associated with rainfall time seriesuncertainty is very complicated and there is a requirement to

    assume a linear and normal distribution within the framework

    of the proposed parameters. The problem of model selection is

    often divided into discover an organization of a models

    parameters that is well-matched such as the network topology,

    e.g. number of patterns, layers, hidden units per layer, that

    results in the best generalization performance. A common

    result is with too many free parameters tend to overfit the

    training data and, thus, show poor generalization performance.

    A model attempting to estimate the value of a random

    variable may have potential access to a wide range of

    measurements regarding the state of the environment. Some of

    these quantities may provide the model with useful informationregarding the random variable, whereas others may not. In the

    context of neural networks, only the useful quantities should be

    used as inputs to a network. A network that receives both

    useful inputs and nuisance inputs will contain too many free

    parameters and, thus, be prone to overfitting the training data

    leading to poor generalization.

    Time series forecasting using Bayesian method:

    application to cumulative rainfall

    C. Rodriguez Rivero, J. Pucheta, M. Herrera, V. Sauchelli and S. Laboret,Member, IEEE

    T

  • 8/11/2019 pdf139.pdf

    2/6

    139 2

    II. BAYESIANAPPROACH

    One of the most key principles of Bayesian technique is to

    construct the posterior probability distributions for all the

    unknown entities in a model, given the data sample. To use the

    model, marginal distributions are constructed for all those

    entities that we are interested in, i.e. the outputs of the ANNs.

    These can be the parameters in parametric models, or the

    predictions in (non-parametric) regression or classificationtasks. Use of the posterior probabilities requires explicit

    definition of the prior probabilities for the parameters. The

    posterior probability for the parameters u in a modelMgiven

    the dataDis, according to the Bayes rule,

    ,)/(

    )/(),/(),/(

    MDP

    MPMDPMDP

    (1)

    where P(/D,M)is the likelihood of the parameters , P(/M)

    is the prior probability of , and P(D/M) is a normalizing

    constant, called evidence of the model M. The termMdenotes

    all the hypotheses and assumptions that are made in defining

    the model. All the results are conditioned on these

    assumptions, and to make this clear we prefer to have the term

    Mexplicitly in the equations. In this notation the normalization

    term P(D/M) is directly understandable as the marginal

    probability of the data, conditional on M, integrated over

    everything the chosen assumption M and prior P(/M)

    comprise

    dpxxpxxp nene )(),/()/( (2)

    when having several models, P(D/M) is the likelihood of the

    model i, which can be used in comparing the probabilities of

    the models, hence the term evidence of the model. A widely

    used Bayesian model choice method between two models is

    based on Bayes factors,P(D/M1)/P(D/M2).The more common

    notation of Bayes formula, with M dropped, more easily causes

    misinterpreting the denominator P(0) as some kind of

    probability of obtaining data D in the studied problem (or prior

    probability of data before the modeling).

    The prior distribution expresses our initial beliefs about

    parameter values before any data is observed. After new data

    D = {(x(1)

    , xe(1)

    ), , (x(n)

    , xe(n)

    )} are observed, the prior

    distribution will be updated to posterior distribution using

    Bayes rule where L(/|D) is likelihood function of unknown

    model parameters to the observed data. In case of independent

    and exchangeable data points, the likelihood function is

    ),/()/(1

    )()(

    n

    i

    ii

    e xxPDP (3)

    where nis the number of data points.

    To predict the new output xe(n+1) for the new input

    x(n+1), predictive distribution is obtained by integrating the

    predictions of the model with respect to the posterior

    distribution of the model parameters

    ,)/(),/(

    ,/()/(

    )1()1(

    )1()1(

    dDPxxD

    xxPDP

    nn

    e

    nn

    e

    (4)

    where is the space of all possible parameters. Note that

    predictive distribution for xe(n+1) is implicitly conditioned on

    hypotheses that hold throughout and to be more explicit as the

    following [51]

    ,),/(),,/(),/( )1()1()1(

    dMDPDxxMDxP nn

    e

    n

    e

    (5)

    whereM refers to the set of hypotheses or assumptions used to

    define the model. Practically, the posterior distribution for

    parameters in Eq. (4) is very complex and with many modes.

    As a result, evaluating the above integral is a difficult task.

    Neal [7] introduced Markov Chain Monte Carlo (MCMC)

    method to perform this kind of difficult integrations. Then the

    MCMC methods have been utilized by other authors.

    III. ARQUITECTUREOFTHEANN

    In Fig. 1 the block diagram of the nonlinear predictionscheme based on a ANN filter is shown. Here, a prediction

    device [15]-[13] is designed such that starting from a given

    sequence {xn} at time n corresponding to a time series it can be

    obtained the best prediction {xe} for the following sequence of

    18 values. Hence, it is proposed a predictor filter with an input

    vector lx, which is obtained by applying the delay operator, Z-1

    ,

    to the sequence {xn}. Then, the filter output will generatexeas

    the next value, that will be equal to the present value xn. So,

    the prediction error at time k can be evaluated as:

    (6)

    which is used for the learning rule to adjust the NN weights .

    Fig. 1. Block diagram of the nonlinear prediction.

    The coefficients of the nonlinear filter are adjusted on-linein the learning process, by considering an online heuristic

    criterion that modifies at each pass of the time series the

    number of patterns, the number of iterations and the length of

    the tapped-delay line, in function of the Hursts value H

    calculated from the time series taking into account the

    Bayesian inference of the output values[9].

    According to the stochastic behavior of the series, H can

    be greater or smaller than 0.5, which means that the series

    tends to present long or short term dependence [14],

    respectively.

    Estimation of

    prediction error

    Z-1 I Error-correctionsignal

    One-step

    prediction

    NN-Based

    Nonlinear Filter

    Input signal

    kxkxke en

  • 8/11/2019 pdf139.pdf

    3/6

    139 3

    Furthermore, the NNs weights are tuned by means of the

    Levenberg-Marquardt rule, which considers the long or short

    term stochastic dependence of the time series measured by the

    Hursts parameter H. The learning rule consists of changing the

    number of patterns, the filters length and the number of

    iterations for each corresponding time series using the Bayesian

    inference in accordance with Hurst parameter H taking into

    account that the series forecasted has the same H as the real

    time series.

    A. Application to ANN predictor

    When a short or long series is being analyzed, it is

    important to make use of the simplest possible models.

    Specifically, the number of unknown parameters must be kept

    at a minimum.

    The gamma distributions have been considered in the

    literature for this purpose. When a Bayesian analysis is

    conducted, inferences about the unknown parameters are

    derived from the posterior distribution. This is a probability

    model which describes the knowledge gained after observing aset of data. The application of the regression problem involving

    the correspond neural network functiony(x,w) and the data set

    consisting of N pairs, input vector lx and targets tn(n=1,.,N)

    Assuming Gaussian noise on the target, the likelihood

    function takes the form:

    ,);(2

    exp2

    ),/(1

    22/

    N

    n

    nn

    N

    twxyMwDP

    (7)

    where is a hyper-parameter representing the inverse of the

    noise variance. We consider in this work a single hidden layer

    of tanh units and a linearoutputs units.

    To complete the Bayesian approach for this work, prior

    information for the network is required. It is proposed to use,

    analogous to penalties terms, the following equation

    ,2

    exp2)(2

    2

    2/2

    w

    wwwP

    N

    (8)

    assuming that the expected scale of the weights is given by w

    set by hand. This was carried out considering that the network

    function f(xn+1,w) is approximately linear with respect to win

    the vicinity of this mode, in fact, the predictive distribution for

    yn+1will be another multivariate Gaussian.

    B. Performance measure of the ANN predictor filter

    In order to test the proposed design procedure of the ANN

    predictor, an experiment with time series obtained from the

    MG solution and cumulative rainfall time series was performed.

    The performance of the filter is evaluated using the Symmetric

    Mean Absolute Percent Error (SMAPE) proposed in the most

    of metric evaluation, defined by

    1002

    1

    1

    n

    t tt

    tt

    S

    FX

    FX

    nSMAPE (9)

    where tis the observation time, nis the size of the test set,sis

    each time series, Xt and Ftare the actual and the forecasted

    time series values at time t respectively. The SMAPE of each

    series s calculates the symmetric absolute error in percent

    between the actualXt and its corresponding forecast value Ft,

    across all observations tof the test set of size nfor each time

    series s.

    IV.

    MAINRESULTS

    A. MG and Santa Francisa Rainfall generations

    The time series are obtained from the solution of the MG

    equation[10]. The MG equation is explained by the time delay

    differential equation defined as:

    ),()(1

    )()( ty

    ty

    tyty

    c

    (10)

    and rainfall time series obtained from Santa Francisca, San

    Bartolom and La Sevillana, from Cordoba, Argentina, with

    parameters shown in Table 1. This collection of coefficients

    was chosen to generate time series MG1 and MG2 whose H

    parameters vary between 0 and 1. The chosen one was selected

    in accordance to its roughness.

    Table 1. Parameters to generate the times series.

    Time Series Type Parameters H

    1 MG1 =1.9 0.22

    2 MG2 =2.1 0.13

    3 Cumulative Rainfall Santa Francisca 0.022

    4 Cumulative Rainfall San Bartolom 0.0132

    5 Cumulative Rainfall La Sevillana 0.259

    B. Set-up of the ANN algorithm

    The initial conditions for the filter and learning algorithm

    are shown in Table 2. Note that the first number of hidden

    neurons and iteration are set in function of the input number.

    These initiatory conditions of the learning algorithm were used

    to forecast the primitive of the time series, whose length is 102

    values.

    TABLE 2. INITIAL CONDITIONS OF THE PARAMETERS

    Variable Initial Conditions

    lx 12

    Ho 7

    it 100

    H 0.5

    C. Time Series Prediction Results

  • 8/11/2019 pdf139.pdf

    4/6

    139 4

    Each time series is composed of the MG solutions and

    cumulative rainfall time series. However, there are three classes

    of data sets: one is the original time series used for the

    algorithm in order to give the forecast, which comprises 102 or

    79 as in La Sevillana rainfall series values. The other one is the

    primitive obtained by integrating the values of original time

    series and the last one is used to compare if the forecast is

    acceptable or not where the 18 last values can be used to

    validate the performance of the prediction system, which 102or 61 values form the data set, and 120 or 79 values constitute

    the Forecasted and the Real ones. A comparison is made

    between this work and earlier presented in [18] ANN H

    dependent predictor filter.

    The Monte Carlo method was used to forecast the next 18

    values from MG time series and rainfall time series. Such

    outcomes are shown from Fig. 2 to Fig. 6.

    0 20 40 60 80 100 120 1400

    0.1

    0.2

    0.3

    0.4

    0.5

    0.6

    0.7

    Time [samples].Real Media = 0.10534. Forescated Mean = 0.1261.

    H = 0.2263. He= 0.17241. l

    x= 20. SMAPE = 1.1693.

    H dependent algorithm.Neural Network based predictor

    Mackey Glass parameters: =1.9,=30, c=10, =100.

    Mean Forescasted

    Data

    Real

    Fig. 2. Results obtained from MG1 time series.

    0 20 40 60 80 100 120 1400

    0.05

    0.1

    0.15

    0.2

    0.25

    0.3

    0.35

    0.4

    Time [samples].Real Media = 0.11226. Forescated Mean = 0.097659.H = 0.13835. H

    e= 0.1472. l

    x= 20. SMAPE = 2.2398.

    H dependent algorithm.Neural Network based predictor

    Mackey Glass parameters: =2.1,=40, c=10, =100.

    Mean Forescasted

    Data

    Real

    Fig. 3. Results obtained from MG2 time series.

    0 20 40 60 80 100 120 1400

    0.1

    0.2

    0.3

    0.4

    0.5

    0.6

    0.7

    0.8

    0.9

    1

    Time [samples].Real Media = 0.20763. Forescated Mean = 0.25411.

    H = 0.022508. He= 0.17071. l

    x= 20. SMAPE = 2.5966.

    H dependent algorithm.Neural Network based predictor

    Lluvia Mensual Acumulado Santa Francisca

    Mean Forescasted

    Data

    Real

    Fig. 4. Results obtained from cumulative rainfall time series

    Santa Francisca, Crdoba, Argentina.

    0 20 40 60 80 100 120 1400

    0.1

    0.2

    0.3

    0.4

    0.5

    0.6

    0.7

    0.8

    0.9

    1

    Time [samples].Real Media = 0.14141. Forescated Mean = 0.20693 .H = 0.12252. H

    e= 0.23774. l

    x= 20. SMAPE = 8.7109.

    H dependent algorithm.Neural Network based predictor

    Lluvia Mensual Acumulado San Bartolom

    Mean Forescasted

    Data

    Real

    Fig. 5. Results obtained from cumulative rainfall time series

    San Bartolom, Crdoba, Argentina.

    0 10 20 30 40 50 60 70 800

    0.1

    0.2

    0.3

    0.4

    0.5

    0.6

    0.7

    0.8

    0.9

    1

    Time [samples].Real Media = 0.25847. Forescated Mean = 0.21356 .

    H = 0.25994. He= 0.027721. l

    x= 20. SMAPE = 4.8701.

    H dependent algorithm.Neural Network based predictor

    Lluvia Mensual Acumulado La Sevillana

    Mean Forescasted

    Data

    Real

    Fig. 6. Results obtained from cumulative rainfall time series

    San Bartolom, Crdoba, Argentina.

  • 8/11/2019 pdf139.pdf

    5/6

    139 5

    D. Comparative Results

    TABLE 3. FIGURES OBTAINED BY THE PROPOSED APPROACH

    Series No. H He

    Real

    Mean

    Mean

    Forecasted SMAPE

    MG1 0.22 0.17 0.105 0.126 1.16

    MG2 0.13 0.14 0.11 0.097 2.23

    Santa

    Francisca 0.022 -0.17 0.207 0.254 2.59

    San

    Bartolom 0.12 0.23 0.141 0.206 8.71

    La Sevillana 0.25 0.027 0.258 0.217 4.87

    The assessments of the obtained results by comparing the

    performance of the proposed filter with the earlier work [18]

    [19]and[20], both are based on ANN.

    Although the difference between filters resides only in the

    adjustment algorithm, the coefficients that each filter has, each

    ones performs different behaviors, in which it can be noted that

    rainfall time series have more roughness than MG solutions, so

    the Bayesian approach applied to the parameter of the ANN

    demonstrate a level improvement, in which an adequate prior

    distribution model was chosen in order for tuning the

    parameters and outputs of the predictor filter.

    V. CONCLUSIONS

    In this work an algorithm to adjust parameters using a

    Bayesian method for cumulative rainfall time series forecasting

    implemented by an ANN-filter was presented. The contributionof this work resides in modeling the number of NN in the filter

    that has an uncertainty number of parameters, which is adjusted

    by Bayesian approach. Thereby, these are considered as

    random variables whose posterior probability distribution is

    inferred from the data.

    The learning rule proposed to adjust the NNs weights is

    based on the Levenberg-Marquardt method. Furthermore, in

    function of the long or short term stochastic dependence of the

    time series, the proposed approach changes the number of

    patterns and iterations using the Bayesian inference in

    accordance with parameter H evaluated of each time series. An

    on-line heuristic adaptive law was proposed to update the

    ANN topology at each time-stage taking into account that theseries forecasted has the same H as the real time series. The

    major result shows that the predictor system has an optimal

    performance from several MG time series and cumulative

    rainfall against the classical predictor filter presented in[20], in

    particular to time series whose H parameter has a high

    roughness of signal, which is evaluated by H, respectively.

    These results show a better performance compared with [19].

    This fact encourages us to apply the proposed approach to

    meteorological time series when the observations are taken

    from a single point.

    REFERENCES

    [1] Taner, M. A. (1993). Tools for statistical inference: methods for the

    exploration of posterior distributions and likelihood functions. New

    York: Springer-Verlag.

    [2] Buntine, W. L., & Weigend, A. S. (1991). Bayesian back-propagation.

    Complex Systems, 05(6), 603.[3] D.J.C. MacKay, A practical Bayesian framework for backpropagation

    networks, Neural Comput. 4 (1992) 448}472.

    [4] Neal, R. M. (1992). Bayesian training of backpropagation networks by

    the hybrid Monte Carlo method. Technical report CRG-TR-92-1,

    Department of Computer Science, University of Toronto.

    [5] Bishop, C. (2006). Pattern Recognition and Machine Learning. Boston:

    Springer.

    [6] MacKay, D. J. C. (1995). Probable networks and plausible predictions -

    a review of practical Bayesian methods for supervised neural networks.

    Network:Computation in Neural Systems, 6 (3), 469-505.

    [7] R.M. Neal, in: Bayesian learning for neural networks, Lecture Notes in

    Statistics, Vol. 118, Springer, New York, 1996.

    [8] Liu, J.N.K.; Lee, R.S.T (1999). Rainfall forecasting from multiple point

    sources using neural networks. In proc. of the International Conference

    on Systems, Man, and Cybernetics, 3, 429-434.

    [9] Mendoza, M. & de Alba, E. (2006). Forecasting an accumulated series

    based on partial accumulation II: A new Bayesian method for short

    series with seasonal patterns, International Journal of Forecasting, Issue

    4, 781-798.

    [10] Mandelbrot, B. B., (1983), The Fractal Geometry of Nature, Freeman,

    San Francisco, CA. 1983.

    [11] Masulli, F., Baratta, D., Cicione, G., Studer, L (2001). Daily Rainfall

    Forecasting using an Ensemble Technique based on Singular Spectrum

    Analysis. In Proceedings of the IEEE International Joint Conference on

    Neural Networks IJCNN 01, 1, 263-268.

    [12] Mozer, M. C (1994). Neural Net Architectures for Temporal Sequence

    Processing. Time Series Predictions: Forecasting the Future and

    Understanding the Past. IEEE International Conference on Image

    Processing, 243-264.

    [13] Pucheta, J., Patio, H., Schugurensky, C., Fullana, R., Kuchen, B.

    (2007a). Optimal Control Based-Neurocontroller to Guide the Crop

    Growth under Perturbations. Dynamics Of Continuous, Discrete And

    Impulsive Systems. Special Volume Advances in Neural Networks-

    Theory and Applications. DCDIS A Supplement, Advances in Neural

    Networks, 14(S1), 618-623.

    [14] Pucheta, J., Patio, H.D., Kuchen, B. (2007). Neural Networks-Based

    Time Series Prediction Using Long and Short Term Dependence in the

    Learning Process. In proc. of the 2007 International Symposium on

    Forecasting, New York, USA.

    [15] Pucheta, J., Patino, D. and Kuchen, B. A Statistically Dependent

    Approach For The Monthly Rainfall Forecast from One Point

    Observations. In Book Series Title: IFIP Advances in Information and

    Communication Technology, Book Title: Computer and Computing

    Technologies in Agriculture II, Volume 2, IFIP International Federation

    for Information Processing Volume 294, Computer and Computing

    Technologies in Agriculture II, Volume 2, eds. D. Li, Z. Chunjiang,

    (Boston: Springer), ISBN: 978-1-4419-0210-8, pp. 787798, Url:

    http://dx.doi.org/10.1007/978-1-4419-0211-5_1, Doi: 10.1007/978-1-

    4419-0211-5_1. (2009).

    [16] Pucheta, J., Patio, H., Schugurensky, C., Fullana, R., Kuchen, B.

    Optimal Control Based-Neurocontroller to Guide the Crop Growth under

    Perturbations. Dynamics Of Continuous, Discrete And Impulsive

    Systems Special Volume Advances in Neural Networks-Theory and

    Applications. DCDIS A Supplement, Advances in Neural Networks,

    Watam Press, Vol. 14(S1), pp. 618623. 2007.

    [17] J.A. Pucheta, C. Schugurensky, R. Fullana, H. Patio and B. Kuchen.

    A Neuro-Dynamic Programming-Based Optimal Controller for Tomato

    Seedling Growth in Greenhouse Systems. Neural Processing letters.

    Editorial Springer Verlag (Springer Netherlands). ISSN 1370-4621

    (Print) 1573-773X (Online) DOI 10.1007/s11063-006-9022-9, Volume

    24, Number 3 / December, 2006, Pages 241-260.

  • 8/11/2019 pdf139.pdf

    6/6

    139 6

    [18] J. Pucheta, M., C. Rodrguez Rivero, M. Herrera, C. Salas, D. Patio

    and B. Kuchen. A Feed-forward Neural Networks-Based Nonlinear

    Autoregressive Model for Forecasting Time Series. Revista

    Computacin y Sistemas, Centro de Investigacin en Computacin-IPN,

    Mxico D.F., Mxico, Computacin y Sistemas Vol. 14 No. 4, pp. 423-

    435 ISSN 1405-5546, 2011.

    http://www.cic.ipn.mx/sitioCIC/images/revista/vol14-4/art07.pdf

    [19] C. Rodrguez Rivero, J. Pucheta, J. Baumgartner, M. Herrera, D. Patio

    y B. Kuchen. A NN-based model for time series forecasting in function

    of energy associated of series, Proc. of the International Conference on

    Applied, Numerical and Computational Mathematics (ICANCM'11),Barcelona, Spain, September 15-17, 2011, ISBN 978-1-61804-030-5,

    Pp. 80-86. (2011).

    [20] C. Rivero Rodrguez, J. Pucheta, J. Baumgartner, H.D. Patio and B.

    Kuchen, An Approach for Time Series Forecasting by simulating

    Stochastic Processes Through Time-Lagged feed-forward neural

    network. The 2010 World Congress in Computer Science, Computer

    Engineering, and Applied computing. Las Vegas, Nevada, USA, July

    12-15, 2010. DMIN10 Proceedings ISBN 1-60132-138-4 CSREA

    Press,p.p 278, (CD ISBN 1-60132-131-7), USA, (2010).

    C. Rodriguez Riveroreceived the Electrical Engineering degree from Faculty

    of Exact, Physical and Natural Sciences at National University of Cordoba,

    Argentina, in 2007, and he is currently pursuing the Ph.D. degree in electrical

    engineering at University of Cordoba, Argentina, under supervision of Dr.

    Pucheta and Grant PRH National Agency for Scientific and Technological

    Promotion (ANPCyT), in the field of automatic control for slow dynamics

    processes, stochastic control, optimization and time series forecast. He joined

    the Mathematics Research Laboratory applied to Control in 2009. IEEE

    member of CSS and CIS.

    J. Pucheta received the Electrical Engineering degree from NationalTechnological University - Cordoba Regional Faculty, Argentina, the M.S. and

    Ph.D. degrees from National University of San Juan, Argentina, in 1999, 2002

    and 2006, respectively. He is currently Professor at the Laboratory of research

    in mathematics applied to control, National University of Cordoba, Argentina.

    His research interests include stochastic and optimal control, time series

    forecast and machine learning. IEEE member.

    M. Herrera received the Electrical Engineering degree in 2007 from Facultyof Exact, Physical and Natural Sciences at National University of Cordoba,

    Argentina. He is Associated Professor and researcher from Faculty of AppliedSciences and Technology at National University of Catamarca. His research

    interests include control systems and automation.

    V. Sauchelli received the Electrical Engineering degree and Ph.D. degreesfrom National University of Cordoba, Major in University Education from

    National Technological University, Crdoba Regional Faculty (UTN, FRC,)

    Argentina, in 1973 and 1997, respectively. He is currently Principal at

    Telecommunication Postgraduate Program, National University of Cordoba,

    and Professor of Control Systems and Signal Processing in electrical and

    electronic engineering. He is also Director of Director of Mathematics Research

    Laboratory applied to Control (LIMAC)and Research Projects at Secretary of

    Sciences and Technology (SECyT), National University of Cordoba. His

    interests include automatic control, neural networks, fractional-order controllers

    and signal processing.

    http://www.cic.ipn.mx/sitioCIC/images/revista/vol14-4/art07.pdfhttp://www.cic.ipn.mx/sitioCIC/images/revista/vol14-4/art07.pdfhttp://www.cic.ipn.mx/sitioCIC/images/revista/vol14-4/art07.pdf