Using the Neural Network Time Series Tool

Using the Neural Network Time Series Tool Tool 1. If nee neede ded, d, op open en th the e Neu Neural ral Ne Netw twork ork St Start art GU GUII wit with h thi this s com comman mand: d: 2.nnstart

3.

Click Time Series Tool to Tool to open ope n the Neural Network Time Series Tool. Tool. !ou !o u can also use the commandntstool."

•

•

Notice that this openin# pane is different than the openin# panes for the other GUIs. This is $ecause ntstoolcan $e used to sol%e three different kinds of time series pro$lems. In the first t&pe of time series pro$lem, &ou would like to predict future %alues of a time series y t " from past %alues of that time series and past %alues of a second time series x  series x t ". ". This form of prediction is called nonlinear autore#ressi%e with e'o#enous e'ternal" input, or N()* see +N()* Network+ nar'net, Network+  nar'net, closeloop"", and can $e written as follows: y t "  f y t   - 1", ..., y t   - d ", x  ", x t - 1", ..., t  t   - d "" "" This model could $e used to predict future %alues of a stock or $ond, $ased on such economic %aria$les as unemplo&ment une mplo&ment rates, G/, etc. It could also $e used for s&stem identification, in which models are de%eloped to represent d&namic s&stems, such as chemical processes, manufacturin# s&stems, ro$otics, aerospace %ehicles, etc. In the second t&pe of time series pro$lem, there is onl& one series in%ol%ed. The future %alues of a time series y t " are predicted onl& from past %alues of that series. This form of prediction is called nonlinear  autore#ressi%e, or N(), and can $e written as follows: y t "  f y t   - 1", ..., y t   - d "" ""

This model could also $e used to predict financial instruments, $ut without the use of a companion series. The third time series pro$lem is similar to the first t&pe, in that two series are in%ol%ed, an input series x t " and an output0tar#et series y t ". ere &ou want to predict %alues of y t " from pre%ious %alues of x t ", $ut without knowled#e of pre%ious %alues of y t ". This input0output model can $e written as follows: y t "  f  x t  - 1", ..., x t  - d "" The N()* model will pro%ide $etter predictions than this input2output model, $ecause it uses the additional information contained in the pre%ious %alues of y t ". owe%er, there ma& $e some applications in which the pre%ious %alues of y t " would not $e a%aila$le. Those are the onl& cases where &ou would want to use the input2output model instead of the N()* model. . 4or this e'ample, select the N()* model and click Next to proceed. •

5.

Click Load Example Data Set in the Select ata window. The Time Series ata Set Chooser window opens. Note Use the Inputs and Targets options in the Select Data window when you need to load data from the MATLAB ® workspace.

6.

7.

Select pH Neutralization Process, and click Import. This returns &ou to the Select ata window. 8. Click Next to open the 9alidation and Test ata window, shown in the followin# fi#ure. The %alidation and test data sets are each set to 15 of the ori#inal data.

;ith these settin#s, the input %ectors and tar#et %ectors will $e randoml& di%ided into three sets as follows: • •

•

7 ) Pre!ict $ne Step Ahea!>7+ /=. viewnets4 /C. 6xs3xis3ais3ts7 * preparetsnets3inputSeries3 893targetSeries4+ /;. "s * netsxs3xis3ais4+ /D. earl"Pre!ictPer&ormance * per&ormnets3ts3"s4 /. earl"Pre!ictPer&ormance * =1. =.

1.11/2

=2.

4rom this fi#ure, &ou can see that the network is identical to the pre%ious open2loop network, e'cept that one dela& has $een remo%ed from each of the tapped dela& lines. The output of the network is then y t  J 1" instead of y t ". This ma& sometimes $e helpful when a network is deplo&ed for certain applications. If the network performance is not satisfactor&, &ou could tr& an& of these approaches: •

• • •

•

•

•

)eset the initial network wei#hts and $iases to new %alues with init and train a#ain see +Initiali=in# ;ei#hts+init"". Increase the num$er of hidden neurons or the num$er of dela&s. Increase the num$er of trainin# %ectors. Increase the num$er of input %alues, if more rele%ant information is a%aila$le. Tr& a different trainin# al#orithm see +Trainin# (l#orithms+". To #et more e'perience in command2line operations, tr& some of these tasks: urin# trainin#, open a plot window such as the error correlation plot", and watch it animate. /lot from the command line with functions such as plotresponse, ploterrcorr and plotper&orm. 4or more information on usin# these functions, see their reference pa#es."  (lso, see the ad%anced script for more options, when trainin# from the command line. Kach time a neural network is trained, can result in a different solution due to different initial wei#ht and $ias %alues and different di%isions of data into

trainin#, %alidation, and test sets. (s a result, different neural networks trained on the same pro$lem can #i%e different outputs for the same input. To ensure that a neural network of #ood accurac& has $een found, retrain se%eral times. There are se%eral other techni?ues for impro%in# upon initial solutions if hi#her accurac& is desired. 4or more information, see Impro%e Neural Network Generali=ation and (%oid %erfittin#.