Home

The winGamma- User Guide - Pages supplied by users

image

Contents

1. 0 000205 56 3 6 1 1 Using the WhatIf and Query options on the LLR model 58 3 6 1 2 A histogram of prediction errors for LLR model 59 3 6 2 Building and testing a neural model 0 00 e eee 59 3 7 Example model construction and testing for DH 34 5000 asc 60 3 7 1 How the prediction quality degrades into the future 60 3 8 Using a prediction file is eX a pex ee ARA ERGO E EX E eg 61 3 8 1 Using a prediction file on Input Output data lesse 61 3 8 2 Using a prediction file on Time Series data lesse 61 3 9 Using the neural networks outside of winGamma 0 0 cee eee 61 3 9 1 The activation function and the sigmoidal 61 3 9 2 NEREIST ooa s epe oo eT ee tC eS ed e ec eae 62 3 9 3 Exporting and using Neural network models in Excel 62 APPENDIX I General Information ooa Cer ERE eR PC SR 64 Shipping list i241 4o RUE SER AQ REESE E E a A R D S RERSAE 64 Hardware requirements amp iss her qx VES SR MER eek wl a ice S 64 Installatioh eielerrisAecepecer cepexsauqbbrevcepksumheqq E pa 64 List of files and directory structure after installation llle 65 Problems reDorbell 365 uL Mr soto oU ust c a Mea NL AL carat ide 66 APPENDIX II Data file formats 02 uo cone e e iren MuR Gb eas TP a ey e DRE Re SN Gees 67 JAMES series data sixes IRSE LU PERIERE CADRE ORA MDC o bog 67 Input OWtput dala nic c
2. Unique Points 10578 10578 0 0 note the SE first increases and l l Unique Points 10578 10578 1 1 1 1 then plateaus Along the plateau a minimum SE occurs at around 3 FEN pmax 20 which from now on Zero Near Neighbours f0 O id we take as the best pmax for Upper 95 Confidence 0 Jo further analysis of this data OU ts Figure 2 21 Shows the M test and we can see that for M 9000 we are beginning to get a stable asymptote From this we infer that around 9000 data points will be required to build a model which will predict with an accuracy about equal to the noise level The result of a Gamma test on pmax 20 near neighbours using the full data set is shown in Table 2 3 4 The winGamma User Guide PERFORMING AN ANALYSIS Version 18 Jan 2002 Figure 2 22 shows the scatter plot and regression line zoomed This is typical of slightly noisy data but the regression line fit is good The desirable empty wedge is the top left corner is not obviously present another indication that the data is somewhat noisy The 3D Histogram in Figure 2 23 shows similar features except that we can see the actual frequency distribution of points in the scatter plot and this shows that strong outliers or empty wedge points although present are relatively infrequent The Angle histogram of Figure 2 24 shows a roll off in frequency as we approach angles close to 7 2 The final Moving Window Gamma Test
3. 3 6 1 1 Using the WhatIf and Query options on the LLR model The Whatlf option allows us to see what happens if we set values for all of the inputs except one and vary the remaining input over some range This is a very useful tool in a variety of contexts For example in a sales and marketing campaign we may be able to answer the question If I spend X on advertising on TV and Y on advertising in newspapers how will the sales of the soft drink vary with the mean day time temperature Similarly the Query option allows a particular selection of all inputs to be queried The use of Predict is discussed in section 3 8 Having analysed the data built and tested a model we can now ask some interesting questions regarding the solar csv data For example using the WhatIf options we can answer the question How does the power output vary when the temperature is fixed at 7 degrees and the Irradiance varies from 0 to 30 The answer is given in Figure 3 6 As expected at a fixed temperature the power output is almost linear with the Irradiance What If P i i 1 i i 1 1 i i 006 18 336 48 666 78 996108 12 132144 162174186 204216228 24 252264 282294 1 156212258324 38 436492548604 66 716772 856912968 1052 1136 122 1304 1388 1472 Input Input Inp Figure 3 6 The variation of output power as Figure 3 7 The variation of output power as Irradiance varies from 0 to 30 and Temperature varies from 1 to 15 and temperature is 7 de
4. Primary Y Series Gamma zl a Gamma zl Overlay Y Series Eerad 1 3455 7 8 9101112131415 1517 18 1920 21 2223 24 2526 27 2829 100 120 140 160 180 200 220 240 260 280 300 320 340 360 380 400 420 440 460 480 S Near Neighbours Unique Data Points Gamma Standard Error Figure 2 13 Increasing near neighbours 3 Figure 2 14 M test pmax 17 on 30 on Sin500 asc Gamma SE Sin500 asc Analysis Manager a Analysis Manager Select output Output 1 7 Select output Output 1 7 Scatter Plot 30 Histogram Angle Histogram Settings Gamma Scatter Plot 3D Histogram Frequenc BREESE ES 88278 2B 0 01 0015 0 02 0 025 delta Figure 2 15 Scatterplot and regression line Figure 2 16 3D Histogram pmax 17 for pmax 17 for Sin500 asc Sin500 asc Analysis Manager E Results Visualiser Select output Output 1 X Select output Output 1 zl Scatter Plot 3D Histogram Angie Fisiogram Settings Cision Chai X Series N Angle Histogram Position in List z Position in List v Gamma 0 0765 0 0760 0 0755 00750 goos t 5 amp 0 0740 x E 2 5 5 G Fy 00735 0 0730 0 0725 0 0720 9 10 11 12 13 14 15 16 17 18 19 20 Position in List 90 80 70 60 50 40 30 20 10 50 5 101520 2530 3540 45 50 5560 65 70758085 zn Figure 2 17 Angle histogram for Sin500 asc Figure 2 18 Moving window Gamma test pmax 17 pmax 17 on 300 p
5. File and then Open Analysis Data Set 1 2 1 Comma separated variable csv files from spreadsheets If the file data is in the csv format e g as exported from Excel on loading the file you will be asked to specify which of the columns are outputs Because a csv file does not indicate which columns are inputs and which are outputs if the file is an Input Output file it is necessary to give this information to winGamma Each column has to be tagged as an input or output column This is done as indicated in Figure 1 1 To change an input default to an output select it with the mouse or cursor keys and press the Enter or Return key or toggle with a double click on the left mouse button 11 The winGamma User Guide GETTING STARTED Version 18 Jan 2002 alia eae Data Transformation s Data Settings Time Series Options Number of inputs per series 5 Number of outputs per series 2 Moving average width o Differences a O Output Use the cursors or mouse to select a row Press return or double click to toggle the highlighted row between input and output Cancel Apply Figure 1 1 Toggle inputs to outputs as Figure 1 2 Selecting the number of inputs required when loading a csv file as and outputs per time series Input Output data For Time Series data specify all columns as inputs As in Figure 1 2 winGamma will then ask you to specify the number of inputs and outputs per series A
6. List of Tables Table 1 1 Gamma test results with pmax 10 for unscaled and scaled solar csv data 15 Table 2 1 The results of a simple Gamma test on the file Ran500 asc for unscaled and scaled data 35 Table 2 2 The Gamma test result pmax 10 for unscaled and scaled data on the file Sin500 asc 37 Table 2 3 The results of the Gamma test pmax 20 for unscaled and scaled data from solar csv 41 Table 2 4 Excel file for multiple time series 0 6 eee I 48 CHAPTER I Getting Started 1 1 Introduction Data or observations can be considered as a spreadsheet of numbers in which the columns are divided into two types input columns and output columns In any row we might wish to determine the values of the outputs when these are not known but the corresponding inputs are known A data model data model is an algorithm constructed from a set of observations for which all inputs and outputs are known which enables us to predict the outputs from a given set of inputs This software is concerned with constructing data models of a particular type 1 1 1 The Purpose of the Software winGamma is a software package which in the first instance estimates the least Mean Squared Error MSError that any smooth data model e g a trained feed forward neural network can achieve on the given data without over training winGamma can be used with multiple column nput Output data files and single or mult
7. This test can also tell us how much data we are likely to need to obtain a model of a given quality Moving Window Gamma test Shows how the estimate for the Gamma statistic using a fixed number of data points varies as we move a fixed length window along the data file This is used to check the stability of the Gamma statistic as we move along a large file Model Identification These options are used to select those inputs which can best be used to predict a selected output some inputs may be noisy or irrelevant The use of model identification techniques is discussed in Chapter II Full Embedding Genetic Algorithm Hill Climbing Sequential Embedding Increasing Embedding Other features Are captioned in Figure 1 7 winGamma Scale unscale or partition the data E META Options View Window Help t Transform data set ate ane T2 BG Data Set Manager oo Partition analysis data set Current data file name Analysis Ma apet so al B Delete selected experiment as Analysis Manager Select Analyse for graphical D vs xx Analyse Graph Model Test Guey whati Predict analyses Experiments l Models Results Settings Current Experiment type E Training Set Analysis ajos Output 1 z i Gamma test Current Experiment number it i m Experiment 1 Analyse Model Row 1 Increasing near neighbours Gradient Standard Error V Raatio Current Experiment results toe MTest WIE EM
8. 19 1 5 1 An Input Output file a Veo D RI eH Gries n 5 QUER aA C REC 19 1 5 1 1 The basic Steps ose AG Re KHER REA AS CAR ERE ERY 19 3 2 A chaoue Time Series cod oy lea sep SPESE ia O SRI S E E EI 20 1 5 2 T The basic Steps kel E EV ESO UT EE a ERES 21 IGA Mear Jnodels ce deep A ote gh eere hk on bole E oy grea ach es ni oe 23 1 7 Exporting results for use by other software llle 23 1 8 Customising the file and project directories llseleleeeeeeese 23 CHAPTER II Performing an analysis iecore ER as oo 04 nib eek Ra 25 2 1 Introduction oysa sia eo takers S E E AE eyed en uua MES EVE 25 2 LlITheusercycle riii unine d wea ene ea RN GP REC a 25 2 2 The Gamma test eio eere he vec eR EF Ch vene SO pee der pede 26 2 3 The Gamma Test analysis graphs 0 0c cece eee eee ee 27 2 3 1 The scatter plot and regression line 0 0002 ee 27 2 3 2 Phe SLM MISLOSTAMN e Dd oor ona DIE oes Pa ete EE MIROR tx Gee 29 2 9 9 The angle histogram 442 etwa oe eave mE Rae a 29 2 4 Increasing near neighbours oo eod M a te vod SR ss 30 2 5 MAGS i est einer Eus pNESREARERSERIMASRUECRSARENSES PARERE ek 31 2 6 Moving Window Gamma test 0 0 cece eee eee eee ee nee 32 2 7 Model identification 2 002 05 40 22bde nee ranerne 32 2At Pull embedding 345 won Gade RES a a a ee A 32 2 12 Geneue Algorithmi 42 bees ars S i a EARS I a GA NA Ree 33 2 Tuo BEC MAM DING S oy S koe CRANE Ed a eV S dO 33 2 7 4 Sequential Embedding 925
9. 99 02 31p 107 677 03 02 99 04 08p 1 228 288 03 02 99 04 20p 968 704 02 17 99 03 32p 35 328 Real data files Data 11 21 98 11 51a 02 04 99 02 35p DIR Solar 430 515 Solar csv 12 11 98 03 37p DIR Sunspot Sun280 asc 03 19 98 07 52p 2 240 04 20 98 02 15p 24 543 Artificially generated test data files TestFiles 01 04 99 04 16 98 03 27 08 03 02 99 09 15 98 09 15 98 09 15 98 10 29 98 04 22 98 SunPairs asc 02 18p lt DIR gt Noise 12 21p 50 183 Ran500 asc 03 04p 20 539 Sin500 asc 12 43p DIR NoNoise 04 58p 1 958 Hen100 asc 04 59p 9 830 Hen500 asc 05 00p 19 666 Hen1000 asc 01 26p 983 909 Hen50000 asc 12 51p 9 617 MGIls500 asc 65 The winGamma User Guide APPENDIX I General Information Version 18 January 2002 10 22 98 05 09p 96 097 MGIls5000 asc 02 24 99 02 44p 205 286 ModSin5000 asc 12 09 98 03 02p 98 966 DH 34 5000 asc Mathematica 3 01 files 12 19 98 04 25p 2 317 DataAnaly m 12 19 98 04 25p 2 812 869 DataAnaly nb 03 02 99 07 13p 8 831 DataGen m 03 02 99 07 13p 946 827 DataGen nb 10 01 98 2 07p 38 062 mathlinkGamma nb 01 28 99 04 36p 253 440 GammatTestProject exe 10 29 98 4 18p 50 773 NetReader nb Problems reported Graphics files saved are in billions of colours and cause out of memory errors when attempts are made to load them into some software including WPCorelV8 and Graphics Workshop 66 APPENDIX II Data file formats All data files are in plain ASCII and have the file name
10. Input Output file for which you have to specify the inputs and outputs Becuase the data is not now recognised as multiple time series data the Iterate Model option will not be available at present this can only be used with a single time series 49 The winGamma User Guide PERFORMING AN ANALYSIS Version 18 Jan 2002 2 15 To scale or not to scale If two input variables are incompatible e g temperature in degrees K and altitude in metres both in semantics and in range then the effects of a change in one of them can completely outweigh the effects of a change in the other To ensure that all variables at least start with an equal chance to contribute to an output prediction it is often helpful to apply a standard normalization In this software the standard normalization is that the mean of each input variable is mapped to zero and the standard deviation to 0 5 In later versions we may include an option for the user to select the standard deviation this can have some advantages in model building The effect of normalizing the data is two fold First since the output is also rescaled this will affect the Gamma statistic in a trivial way it will divide it by the square of the new output range The Vratio however will not change due to this effect Second rescaling the inputs can change the near neighbourhood relationships and hence possibly change the associated Gamma value We can detect if this happens as it will also cause Vratio t
11. also sometimes interesting to observe when the Gamma statistic is at a local minimum and the Gradient is at a local maximum as the number of near neighbours varies This criterion seems to be sensitive to noise on different scales of distance in input space 2 5 M test This test is used to show how the Gamma statistic and the other results returned by the Gamma test estimate varies as more data is used to compute it Eventually if enough data is used the Gamma statistic should asymptote to the true noise variance on the output for which it has been computed The M test can also tell us how much data we are likely to need to obtain a model of a given quality in the sense of predicting with a MSError around the noise level In Figure 2 5 we see that in this sense a perfectly adequate model can be built using anywhere from 150 200 data points since the variance of the Gamma statistic after this stage is relatively small compared with its actual value Results Visualiser Select output Dutput 1 7 Custom Chart X Series Unique Data Pc i Unique Data Points v Gamma Primary Y Series Gamma Overlay Y Series mro 200 250 300 Unique Data Points Gamma Figure 2 5 M test graph for Sin500 asc Note the relatively stable asymptote We call these the Terry points after John Terry who first observed the phenomenon 3l The winGamma User Guide PERFORMING AN ANALYSIS Version 18 Jan 2002 Of course using more
12. and that we can make accurate short term predictions but that long term prediction s pus becomes exponentially more difficult Figure 3 11 How the Gamma statistic varies against the number of steps ahead for DH 34 5000 asc 3 8 Using a prediction file Building and testing models when you know the outputs for a corresponding set of inputs is quite interesting but it is purely an academic exercise Sooner or later you will want to make predictions that matter and where the outcomes are not known Perhaps from some large quantity of input data To accomplish this it is first necessary to have a prediction file i e the input data is placed in a file without the corresponding inputs 3 8 1 Using a prediction file on Input Output data Load the data file EXAMPLE NEEDED HERE 3 8 2 Using a prediction file on Time Series data Load the data file EXAMPLE NEEDED HERE 3 9 Using the neural networks outside of winGamma If the neural models are used outside of winGamma i e in other software it is necessary to know some technical details of the implementation 3 9 1 The activation function and the sigmoidal The activation function used by the neural networks is 61 The winGamma User Guide BUILDING AND TESTING A MODEL Version 18 Jan 2002 n act x D WX j l j i where Wij is the weight of the connection from unit j to unit 1 and xj is the output of unit j The sigmoidal used by each neural node as an output function i
13. data we can actually often progressively improve the model this can easily be checked by building a local linear regression model and using the WhatIf option to recover a quite good approximation of the original sine curve but it is not necessarily helpful to have an extremely accurate model if the output data we are comparing it with is subject to large amounts of noise 2 6 Moving Window Gamma test The Moving Window Gamma test shows how the estimate for the Gamma statistic and other relevant results returned by the Gamma test using a fixed number of data points varies as we move along the data file It gives some indication of how stable the Gamma statistic is when estimated for different subsets of the data all having the same size The remaining sections deal with model identification i e in this context the best choice of inputs for predicting a given output 2 7 Model identification 2 7 Full embedding An embedding is a selection of inputs chosen from all the possible inputs In winGamma an embedding is designated by a string of 1 s and 0 s called a mask Thus if there five inputs the mask 10111 indicates that all inputs are to be used are to be used in the embedding except the second A full embedding tries every combination of inputs to determine which combination yields the smallest absolute Gamma value It returns the number of results requested If there are m scalar inputs then there are 2 1 possible embeddings th
14. progressively increase the numbers of hidden units in the two layers 1 winG amma Analysis Manager Di Xx File Edit Transform Options View Window Help 8 x B ei mi New Delete Analyse Gem Mode Test BUE WRIT Predict Experiments Models Settings Real Time Evaluation Local Linear Regression Models gt Local linear regression Dynamic local linear regression E e Neural Networks E e Two layer backpropagation neural ET Conjugate gradient neural network BFGS neural network Training Mean Squared Error 8 10 12 14 16 18 20 22 24 26 28 30 32 34 Cycles MSE Target MSE 4 Initialising Weights Figure 3 3 The Analysis Manager during backpropagation training Two layer backpropagation also requires The initial learning rate must be positive This controls the initial step size in weight adjustment Momentum constant must be positive This controls the extent to which the size and direction of the current step in weight space is influenced by the size and direction of the previous step Setting this parameter to zero means there is no momentum term in the weight adjustment at each step Regularisation constant must be positive This limits the size of weights A zero here corresponds to no restriction on weight magnitude These options are configured using the set up menu shown in Figure 3 1 There is a second tab which allows the user to specify the maxim
15. sunspot data bia Mata tx eo A acu xad dar dc apio du dac bab a ex adu iq Od ea hp SACRA SU OS CU Para 47 Figure 2 37 A test of the LLR model on the data set SunPairs asc blue predicted green actual Ied efroby os s Vos Uva ivt veut ie tel cree 47 Figure 3 1 The first set up menu for two layer backpropagation 54 Figure 3 2 The second set up menu for two layer backpropagation 54 Figure 3 3 The Analysis Manager during backpropagation training 55 Figure 3 4 Selecting a proportion of the data for testing 0 0 0 02 ee eee 57 Figure 3 5 Result of LLR test for solar csv lseeeeeee e 57 Figure 3 6 The variation of output power as Irradiance varies from 0 to 30 and temperature is TIO BIGBS puta Sere LOVAL OS et HUS Ql ar dolo E a ade Gummi g eH QUO gen etree SRG lap ose 58 Figure 3 7 The variation of output power as Temperature varies from 1 to 15 and Irradiance is IU Sees Oe Lu edt ARES Sits heal s QAM UD ale dose uta d Rea 58 Figure 3 8 An error histogram for the LLR test 0 0 0 0 ee eee eee eee 59 Figure 3 9 A topographic plot of the solar csv data 0 fcc eee eee 59 Figure 3 10 A test of the LLR model on the data set DH 34 5000 asc blue predicted green actual Ted erFok ous Su ees ise vM veo Go uas d iu e e DAT oS M MOX 60 Figure 3 11 How the Gamma statistic varies against the number of steps ahead for DH 34 5000 856 5 Lo WS E ERU RAS M ERU C xis 61
16. target MSError This is useful in the event that the partition of the data for training and testing has been altered Clicking Recalc will cause a new Gamma statistic for that part of the data selected as training data to be calculated and hence set a new target MSError for training Alternatively the user can set any target MSError However if the target MSError is much less than the Gamma statistic on the training data then 1 the network may end up being overtrained resulting in poor predictions or ii the training algorithm may never be able to reach the possibly unrealistic target MSError User settable options For each of the neural training algorithms we shall need to specify the number of hidden units Thus each neural network option needs The number of units in the two hidden layers default 5 5 The number of units required to achieve a good model will depend on the complexity of the unknown function we are trying to approximate Unfortunately here there are few general rules to guide us One useful guide is that if the Gradient value returned by the Gamma test is large then the unknown function has regions of high curvature and we shall require more hidden units to approximate it 54 The winGamma User Guide BUILDING AND TESTING A MODEL Version 18 Jan 2002 accurately The best approach is to try to train using relatively few units the defaults are set quite low and if training fails to converge to the target MSError
17. the initial analysis and click OK 5 Select Gamma test from the Experiments Manager and then click on New Results Visualiser output output 1 z Chart b Position in List Position in List v Gamma Position in List Figure 1 13 The result of an Increasing Embedding on Hen500 asc with a maximum of 10 inputs using 10 nearest neighbours Results Visualiser Select output output 1 z i Unique Data Pc Y Unique Data Points v Gamma Primary Y Series i KEA 14 pu Gamma g Overlay Y Series None m 20 40 60 801 00 120140160 180200 220240 260 280300320 340360 380400420 440460 480 Unique Data Points Figure 1 14 The result of an M test on Hen500 asc with 2 inputs using 10 nearest neighbours The fact that this is so is by no means obvious Itis a consequence of a fairly deep theorem due originally to Takens 1981 21 The winGamma User Guide GETTING STARTED Version 18 Jan 2002 Initial results The initial result gives a Gamma statistic of 0 117143 and a Vratio of 0 185337 which is not very encouraging However the real reason for this is that most of the inputs we have selected for the model are irrelevant or not very helpful 6 Nextinthe Experiments Manager under Model Identification highlight Increasing Embedding and then click on New Leave the number of near neighbours set at 10 and click on Execute What this experiment does is to compute the Gamma statistic for a succ
18. the angle histogram in Figure 2 11 Finally the Moving Window Gamma test using a window size of 300 in steps of 10 in Figure 2 12 consistently shows a Gamma statistic between 0 29 and 0 38 These results together indicate that there is no point in going on and trying to produce a smooth model for this data 35 The winGamma User Guide PERFORMING AN ANALYSIS Version 18 Jan 2002 Results Visualiser g Results Visualiser Select output Output 1 zl Select output Output 1 zl Custom Chart X Series Near Neighbour zl d Near Neighbours v Gamma and Standard Error Unique Data Pe zl Unique Data Points v Gamma Primary Y Series 3 A i Primary Y Series me Gamma z hr Gamma z Overlay Y Series ade Bl Overlay Y Series a 10143 piepuets 345678 9 1011121314 15 1617 18 19 20 21 2223 24 2526 27 28 29 30 60 80 100120 140160180 200220 240 260 280 300 320 340 360 380 400 420 440 460 480 500 Near Neighbours Unique Data Points Gamma Standard Error Figure 2 7 Increasing near neighbours 3 30 Figure 2 8 M test pmax 10 on on Ran500 asc Gamma SE Ran500 asc Analysis Manager E Analysis Manager Select output Output 1 Select output Output 1 R Gamma Scatter Plot 3D Histogram Frequency Hox B eB aS du tel fies Beets DT a HUN ee Jail 005 01 015 02 025 03 035 04 045 05 06 085 07 075 delta Figure 2 9 Scatterplot and regression line Figure 2 10 3D Histogram pmax 10 fo
19. thumb for this situation but in practice itis not a major problem to optimise the choice of k using a test set There is also an option whether or not to include a constant term in the locally linear model In general it is better to include this term The final user settable parameter is equivalent to a local principal components threshold filter on the eigenvectors of the local linear model We are tying to predict along the tangent plane of the local flow and eigenvectors corresponding to relatively small eigenvalues probably represent noise and lie outside the tangent plane The threshold decides which eigenvectors we should ignore Setting itlow or zero will essentially include all eigenvectors in the local model the default is around 10 Raising this threshold will filter out more and more eigenvectors For noisy time series one often finds that 0 001 gives quite good results Again the best approach is to experiment on a test set 3 3 Dynamic local linear regression This option is mainly designed for time series analysis It is basically identical to LLR with the additional feature that as new data is seen for the first time it is incorporated into the model You can see the effect of this by starting the model with very little training data and running a test on a large amount of data As new test data is encountered but after the attempt at prediction of course dynamic LLR will make steadily better predictions This is interesting to obs
20. 4654 cou zv bue Eee 33 2 7 5 Increasing Embedding 2252255 o9 40d Wises hele yee sete 34 2 12 Analysing Input Output datas 6 04 nicks ere ar ADR pr d Le 34 2 12 1 The Ran500 asc data lecce RR A 34 2 12 2 The Sin500 asc data oeste e e Le eek RI RR ER RUE 37 2112 3 LHe solar esv daldos oerte dorna ERES NE PESO eR HEN 39 2 13 Analysing Time Series dala amp us ost uA ee Nard See ee 42 2 13 1 The DH 34 5000 asc data Delayed Henon Map 42 2 13 2 The FTSE weekly closing price data lessen 44 2 13 3 The sunspot data ss essc esas Kidda gant dbs idawkiwreaaeheees4 46 2 14 Handling multiple time series 2 0 0 0 0 eee eee ee eee 48 2 15 T scale Of dot 10 Stale orainn Lagos Ree HY dh tanta Beg 50 216 PROVE OS a 2 scien x eal obec NPS BGA ae abu aue doi oak PO dtp dE 50 CHAPTER III Building and testing a model 1 0 0 cece eee 52 S TL THEOOUCUOIE 5 beo talc o oe eG hes ta atom sage obe tate dex gee des 52 3 2 Local linear regression sued le 4p Rede E 4 bab ek des EE SA ee RE UE 52 3 3 Dynamic local linear regression 2241 e c V ES x ee kaw QE DM REY RE MA 33 3 3 Two layer back propagation sseeeeeeeeee ee 54 3 4 Conjugate gradient descent 2 iowa acy ath eden eens PI ER aa Er AERA Oe aS 56 ao BEGGS neural NetWork ick eevee ea Mp Cc C A Md sS LE 56 3 6 Example model construction and testing for solar asc cece eee 56 3 6 1 Building and testing a LLR model
21. EE n 256531 2 32865 o 00078571 Moving window gamma test Model Identification be a Ful embedding b Genetic algorithm Hill climbing Figure 1 7 The Analysis and Data Set Manager windows after performing the initial experiment 18 The winGamma User Guide GETTING STARTED Version 18 Jan 2002 1 5 Two simple examples In this section we further illustrate the use of winGamma using two test files provided with the software 1 5 1 An Input Output file The data for the file Sin500 asc was created via the Mathematica file DataGen nb using the function y Sin x and then adding uniformly distributed noise with a theoretical variance of 0 075 to the y values A point plot of the data is shown in Figure 1 8 1 5 1 1 The basic steps Load the data file and run a simple Gamma test with the number of near neighbours set at 10 default as described in section 1 3 Do not scale the data Note that we do not need to specify the number of inputs and outputs because this file is in standard format Figure 1 8 The noisy sine data The Gamma statistic in the Results window is 0 07355 which is quite close to the theoretical noise variance The Vratio of 0 12762 suggests that we will not be able to predict the value of an output very accurately which in view of the data plot in Figure 1 8 is not too surprising The SE is 0 0037651 which indicates a fair degree of reliability in t
22. ORMING AN ANALYSIS Version 18 Jan 2002 Normal distribution see Figure 2 33 If on the other hand there are clear underlying dynamics in the data then the histogram often shows a bimodal or multimodal distribution see Figure 2 36 2 7 2 Genetic Algorithm This option searches the space of all masks using a Genetic Algorithm GA to find good embeddings The parameters which can be used to control this search are default values of parameters are given in brackets Population Size 100 The size of the population of masks being used throughout the search Mutation Rate 0 01 The probability that an individual bit will be mutated during the reproduction process Crossover Rate 0 5 The chance of inserting a random length run of bits from a parent mask to a child mask i e the probability that a crossover event occurs during the reproduction process Gradient Fitness 0 1 The weighting in the GA fitness function for masks giving a low gradient in the Gamma Test Increasing this weighting will place more emphasis on the relative simplicity of the modelling function Intercept Fitness 0 8 The weighting in the GA fitness function for masks with a low absolute value of the Gamma statistic Increasing this weighting will place more emphasis on the model accuracy Length Fitness 0 1 The weighting in the GA fitness function for masks with a given number of 1 s Increasing this weighting will encourage the selection of masks with f
23. SE 0 00236 represents a good fit for the regression line Vratio The Vratio is defined as Gamma Var output It thus represents a standardised measure of the Gamma statistic and enables a judgement to be formed independently of the output range as to how well the output can be modelled by a smooth function In comparing different outputs or outputs from different data sets the Vratio is a good number to study because it is independent of the output range A Vratio close to zero indicates a high degree of predictability by a smooth model of the particular output If the Vratio is close to one the output is equivalent to random noise as far as a smooth model is concerned In this case Vratio 0 00076 indicates low noise data which we should be able to model quite accurately Near Neighbours number of pmax This is the one user settable parameter in the Gamma test When estimating the Gamma statistic pmax should be selected in relation to the size of the data set For large data sets in the interests of getting a more accurate Gamma statistic we can afford to take the number of near neighbours somewhat larger this depends on a number of factors discussed in Chapter II In general in a Gamma test experiment we should keep the number of near neighbours less than 30 Usually 10 20 is a good choice Start This indicates the row identifier for the first vector selected Unique Points 16 The winGamma User Guide GETTING STARTED Versio
24. Set the number of inputs to 20 the number of outputs initially to one Now perform a simple Gamma test on the full data set which gives 4985 I O pairs with 20 inputs to get an initial idea If the data set is very large use a subset of the data for initial experiments Results This gives an initial Gamma statistic of 0 0042614 and a Vratio of 0 007481 The SE for this result is 0 00125 These initial results are encouraging 2 Next run an ncreasing Embedding test to determine a likely embedding dimension Results If we zoom in on the resulting graph we see Figure 2 28 and infer that a good model is likely to be obtained with 4 or 5 previous values The Gamma statistic for 4 is 0 00019635 for 5 it is 0 0002997 but the lowest value of allis for 7 past values which gives 1 5E 7 Results Visualiser Select output Output 1 Y f Custom Char Position in List v Gamma 4 Position in List Figure 2 28 The result of an Increasing Embedding for the delayed Henon map These very low values suggest that the time series is consists of very low noise or noise free data Examination of the scatter plot and associated graphics supports this view 3 We next Transform the data set to reset the maximum number of inputs to 8 4 We next run a M test to check the stability of the Gamma statistic If the M test produces a stable asymptote we can decide if we really have enough data to support these conclusions A reasonab
25. The winGamma User Guide Abstract This document is the user guide for the winGamma software and is updated to version 1 97 winGamma is a state of the art non linear analysis and modelling tool Keywords Smooth model data analysis prediction Gamma test Feature selection Noise estimation Communication regarding this document should be directed to Antonia J Jones DEPARTMENT OF COMPUTER SCIENCE UNIVERSITY OF WALES CARDIFF PO BOX 916 Cardiff CF24 3XF Telephone 444 292 087 4812 Telefax 44 292 087 4598 Or by email to Peter J Durrant at p j durrant cs cf ac uk Copyright University of Wales Cardiff 1998 2001 About winGamma 9 NOTICE This program is experimental and should be used with caution All such use is at your own risk To the extent permitted by applicable laws all warranties including any express or implied warranties of merchantability or fitness for a particular purpose are hereby excluded The authors and distributors of this software disclaim all liability for direct indirect consequential or other damages in any way resulting from this software This program is protected by copyright You may not copy this program or accompanying documentation without the express written permission of the copyright holder You may not modify this program Copyright O University of Wales Cardiff 1998 2001 Acknowledgements Thanks are due to many people who have contribute
26. aders Save the new file as a ASC DOS text file with the filename suffix csv Edit the file name after saving if the text editor insists on putting an extra txt on the suffix 48 The winGamma User Guide PERFORMING AN ANALYSIS Version 18 Jan 2002 Step 4 Decide the maximum lag you are likely to need in analysis and modelling For example if you think the maximum lag likely to be needed in two months and the data is sampled weekly then the Number of Inputs to be set in winGamma will be 8 Load the csv file into winGamma which because the numeric entries are separated by spaces will treat the file as a multiple time series Selecting the number of inputs as 8 and assuming that you wish to produce a one week ahead forecast select the Number of Outputs as 1 Do not at this stage normalise scale the data this can easily be done later but cannot be undone if it is done at this stage You also have to decide at this stage whether you want to include a running window average for each time series as an input and whether you wish to include successive differences as possible inputs Without the running window average or successive differences the data loaded into winGamma will now be ordered into the following columns TS1 t 8 TS1 t 1 TS2 t 8 TS2t 1 TSm t 8 Tsm t 1 Target t 8 Target t 1 TS1 t TSm t Target t where the outputs at time t have been underlined This is a rather tricky mani
27. al scale This can illustrate more clearly the wedge shaped area It can also be used to quickly ascertain the distribution of outliers We shall call point pairs with large 6 each is a long way from its nearest neighbour and large y the y values of close inputs are far apart strong outliers and techniques for identifying and eliminating such points will be discussed in a later version of this document 2 3 3 The angle histogram To help to further analyse the situation the software also produces an angle histogram as for example in Figure 2 11 for each point in the scatter plot we imagine joining the gamma intercept on the vertical axis of the regression line plot to the scatter point The angle the resulting line makes with the positive horizontal axis is then computed This angle lies between 7 2 1 2 A histogram of the resulting angles is then displayed The feature to look for in this histogram is the frequency of angles close to the right hand end i e close to 7 2 If there are no points close to 7 2 90 degrees this is 29 The winGamma User Guide PERFORMING AN ANALYSIS Version 18 Jan 2002 a good indicator for smooth modelling If there are many points close to 7 2 this is a very bad sign The importance of the distribution close to 7 2 in the angle histogram is another way to visualise the upper left hand wedge of the scatter plot The remaining types of Experiment which can be performed are described in the follo
28. ar embedding is sufficiently high to make the outcome of a full embedding search itself extremely unreliable The resulting very low Gamma statistic of around 0 007 is an artifact of the statistics of the situation with over a million embeddings to search we are quite likely to find one with a very small Gamma The associated SE is 517 This clearly illustrates that a low Gamma statistic on a single data set is not enough to ensure a good model we need to be sure that the SE is acceptable and that an M test illustrates the estimate has stabilised In reality using the time series alone we are lucky to predict the weekly closing FTSE price to within a standard deviation of 80 i e the true Gamma statistic is around 6400 There is a further complication in that we have no real reason to suppose either that the underlying system is describable by a smooth dynamic model or that if so the dynamical system is constant Indeed towards the end of the 10 year period it is noticeable that the local variance of both the system behaviour and as we shall see in Chapter IIT of the errors of predictions increase From this we 45 The winGamma User Guide PERFORMING AN ANALYSIS Version 18 Jan 2002 conclude that either the dynamical system is itself varying or at the very least the constant noise variance assumption is suspect 2 13 3 The sunspot data nmp Activity 1700 1373 Figure 2 34 Plot of the sunspots data file Sun280 asc The data used i
29. are empty because it is assumed that the outputs are unknown The use of prediction files is discussed in section 3 10 In general data files can be divided into two main categories input output files and time series files Creating data files using Excel If data is prepared in a spreadsheet it can be exported to winGamma in the csv format Make sure that the numbers exported are in pure decimal format At present winGamma may read numbers in the xEy format incorrectly When a csv file is loaded the user will be automatically prompted to nominate particular columns as inputs or outputs by selecting with the mouse or using up down cursor keys and the Enter or Return key The mouse may also be used to select then double clicking will change an input to an output an vice versa APPENDIX V Definitions Model A smooth data model is a differentiable function from inputs x x Xm to each output y It is assumed that the data can be represented by an unknown model f so that y dax where r is a stochastic variable which represents noise Gamma test An algorithm to estimate the variance of the noise Var r associated with a particular output Not to be confused with the variance of the output Gamma statistic Often referred to as a Gamma value It is the vertical intercept of the regression line plot and represents our best estimate for Var r Embedding A selection of past values of a time series used to predict the curren
30. art of a project with funding from the European Commission THERMIE Programme and the UK Department of Trade and Industry 39 The winGamma User Guide Results Visualiser Select output Output 1 zl Custom Chart X Series Near Neighbour Primary Y Series Gamma zl Near Neighbours v Gamma and Standard Error Overlay Y Series femme 4 6 8 1012 14 16 18 20 22 24 26 28 30 32 34 36 38 40 42 44 46 48 50 Near Neighbours Gamma Standard Error Figure 2 20 Increasing near neighbours 3 50 on solar csv Gamma SE Analysis Manager Select output Output 1 X Scaiter Fiol 3D Histogram Angle Histogram Settings d Gamma Scatter Plot 001 002 003 004 005 O06 007 008 009 Of Off 012 O13 014 04 delta Figure 2 22 Scatterplot and regression line zoomed pmax 20 for solar csv Analysis Manager Select output Output 1 X Scatter Plot 3D Histogram Argie Histogtan Settings Angle Histogram 30 80 70 60 50 40 30 20 10 50 5 101520253035 4045 50 55606570 758085 Angle Figure 2 24 Angle histogram for solar csv pmax 20 PERFORMING AN ANALYSIS Version 18 Jan 2002 Results Visualiser Select output Output 1 zl Unique Data Pc v x Primary Y Series Gamma zl Overlay Y Series ic Unique Data Points v Gamma Ss 1 000 2000 3000 4 000 5000 6000 7 000 8000 9 00 10 000 Unique Data Points Gamma Figure 2 21 M test pmax 20 Randomise
31. art of the input space By spitting up the input space and building a different model for each part a vast improvement on modelling capability was obtained 28 The winGamma User Guide PERFORMING AN ANALYSIS Version 18 Jan 2002 It is interesting to note that by taking the number of near neighbours pmax much larger than is necessary or desirable for the Gamma test the scatter plot can also reveal periodicities on different scales present in the data although for large pmax the resulting Gamma statistic estimate will be essentially meaningless Consider for example the data provided in ModSin5000 asc This is a 1 Input 1 Output file derived from sampling the graph in Figure 2 2 Figure 2 2 Modulated sine curve used to Figure 2 3 Scatter plot with pmax 100 for generate the Input Output file ModSin5000 asc ModSin5000 asc The scatterplot with pmax 100 is shown in Figure 2 3 This illustrates both levels of periodicity and also shows why to get an accurate Gamma statistic we should take pmax fairly small 2 3 2 The 3D histogram This is just another way of viewing the scatter plot The software can also display the scatter plot as a 3D histogram as for example in Figure 2 10 which can be rotated and examined from different viewpoints Click the left and right pointing red arrows to rotate the viewpoint Default is to display frequency values linearly on the vertical axis but there is also an option for a logarithmic vertic
32. ax and the length of the data M If run times are just too long then the Genetic Algorithm GA can be used with Hill climbing and a Sequential embedding embedding to do a small search around the candidates offered by the GA How should I choose the optimal number of near neighbours pmax in the Gamma test See section 2 4 of the manual How should I choose the optimal number of nearest neighbours k in Local linear regression By experiment with a test set The winGamma User Guide APPENDIX VI Frequently asked questions Version 18 January 2002 What is the best Gamma and what does it mean The best Gamma in the context of a Gamma test is the closest approximation to the asymptotic Gamma statistic which should approach the true noise variance The best Gamma in the context where we have a number of Gamma estimates for different selections of inputs essentially different models assuming these estimates are accurate is the Gamma statistic closest to zero because that suggests the model which should have smallest MSError when predicting outputs from inputs not used in the model construction process Note that if the true noise variance is actually zero and the data is of arbitrarily high precision there is no limit to how accurately we can model the unknown function provided only that we have more and more data In most real life situations there is a positive noise variance remaining even after optimising the s
33. b This shows how to load and use the file GammaTestProject exe and gives examples of each function that can be called NetReader nb This notebook can read in any neural network created and saved by winGamma The program identifies the network type and can then run the network There may be very small differences in the results owing to the fact that this notebook uses a pure form of the sigmoidal function whereas winGamma uses a fine grained discrete lookup table for speed in training APPENDIX IV Generating test files Generating your own data files Data files may be generated using a wide variety of software tools AII data files used by winGamma are in plain ASCII format One convenient method of generating data files is to use Excel to manipulate your data into the required rows and columns and then save the data in csv format Another convenient method for creating data files is to use Mathematica winGamma is supplied along with a number of useful Mathematica programs for generating manipulating and saving data in the correct formats These are described in Appendix III Data is generally divided into four types analysis training testing and prediction Prediction files are different in that they contain no output values but otherwise use the same formatting conventions We use prediction files when we genuinely do not know the corresponding output values and want to generate predictions For the prediction file the output fields
34. d 2 repetitions on solar csv Analysis Manager Select output Output 1 X gt 3D Histogram Figure 2 23 3D Histogram pmax 20 for solar csv Results Visualiser Select output Output 1 zl Custom Chart X Series Position in List z Primary Y Series Gamma z Overlay Y Series Position in List v Gamma 3 4 5 8 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 Position in List Gamma Figure 2 25 Moving window Gamma test pmax 20 on 8400 points in steps of 100 from solar csv The winGamma User Guide PERFORMING AN ANALYSIS Version 18 Jan 2002 There are some other factors of interest in this particular situation As it happens the sensor which measures Irradiance is on the roof of the building whereas the solar array is in a nearby location The solar array is shaded at certain times of day by a chimney and a nearby building and this shading is not measured by the Irradiance sensor Since the shading is obviously a function of the time of day and the time of year this has the effect of introducing smooth non linearities into the situation which would be extremely hard to model analytically One could imagine including the time of day and year into the data set and then building a different and much more accurate model By examining the difference between the two models we could actually quantify the effect of shading without having an analytic model This would be a good example of one type of application of
35. d probably build a pretty good model using only around 100 points However if we want to be sure then we should choose around 280 points because from this point onwards the variations in the M test graph are very small 280 points gives a Gamma statistic of 0 001054 and a Vratio of 0 001017 8 Highlight the result in question in row 28 and click Analyse The scatter plot and regression line is shown in Figure 1 15 Handy Tip By left clicking and dragging the mouse down and to the right we can zoom in on any selected part of these graphs as shown in Figure 1 15 We can also move the contents up down and left right by right clicking and dragging To restore the original view simply left click and drag the mouse up and to the right Figure 1 15 The scatter plot for 280 test points on Hen500 asc with 2 inputs using 10 nearest neighbours 22 The winGamma User Guide GETTING STARTED Version 18 Jan 2002 It is interesting to see the remarkable difference between this scatter plot and corresponding scatter plot for the noisy sin data we used in section 1 5 Here in Figure 1 15 we cannot fail to observe the almost empty wedge in the top left hand corner of the plot We shall see in Chapter II that such a feature in the scatter plot is strongly indicative of a noise free smoothly determined process This observation is reinforced by the very small Vratio 9 Finally examine and compare the other graphs produced by the Analy
36. d to the development of winGamma In particular we should like to mention A albj rn Stef nsson whose conviction that MSError could be determined without trial and error sparked the whole idea in 1995 Antonia J Jones and Nenad Kon ar who slaved away to turn an idea into a useable prototype Alban Tsui who explored the outer reaches of the technique using incredibly complicated surfaces and a long double precision version of the Gamma test Steve Margetts who generated the original version of the amazingly tricky tree code in a matter of days and who put an unreasonable amount of effort into the algorithms that are the heart of winGamma Peter Durrant who put far more into the front end development than any of us had a right to expect and who gave the product the look and feel it now has The collaboration between Steve and Pete has been critical McCann Erickson UK Who generously funded a Research Studentship without which winGamma would not now be available Nick Fiddean the HoD of Computer Science at UWC whose unfailing support made the job possible Tina Thomas who put up with an invasion of computer scientists at the research farm over several years and cared for us all so well Howard James who saw the potential and used his business expertise to help turn winGamma into a product Nick Bourne of the RACD division at UWC who has steered us enthusiastically through the legalities that computer scientists blithely tend to ign
37. e embedding where no inputs are chosen can obviously be omitted If m 20 this is around one million To do a full embedding we therefore have to perform one million or so Gamma tests which is fairly time consuming although it can be done in about a week on a fast PC Even if m is sufficiently small to make this practical say m lt 20 before we perform a full embedding assuming say m gt 10 we should ask if we have sufficient data to justify it because looking at around one million Gamma values the differences between many of them will probably be quite small and so we should ask if our estimates of the Gamma values are accurate enough to be able to make these distinctions Whether or not the estimates are sufficiently accurate to choose the absolutely best embedding will mainly depend on how much data is available It practice the best few embeddings will usually have little to choose between them Because a full embedding on a large number of inputs is often pointless or impractical winGamma offers a number of excellent heuristic methods to find a good embedding and these are described in the following sections A useful feature associated with a full embedding or GA search is the Embedding Histogram which shows the frequency of embeddings with a specific Gamma statistic If the choice of embedding is largely determined by statistical variations in the data this histogram tends to have a Gaussian or 32 The winGamma User Guide PERF
38. e model type set at Local Linear Regression set the number of nearest neighbours at 20 leave the Add constant box checked and leave the Define local flow threshold option at 1E 6 Then click on Build 5 When the Test Query WhatIf and Predict buttons become active in the Analysis Manager the model is built and ready to be used Click on Test 6 In the Select proportion of data set for model testing window set the range of test data to 8400 10578 as shown in Figure 3 4 Select proportion of data set for model testing 10578 End Figure 3 4 Selecting a proportion of the data for testing Model Tester Select Output to view JIEN a0 Mean Squared Error 0 011959 Chart Data Model Tester 800 1 000 1 200 1 400 1 500 Actual Predicted Error Figure 3 5 Result of LLR test for solar csv 57 The winGamma User Guide BUILDING AND TESTING A MODEL Version 18 Jan 2002 We have used the points from 8400 10578 for testing A sequence of such experiments for the number of near neighbours k 10 15 20 25 shows that k 20 seems to give the smallest MSError 0 011959 on the test set This is somewhat better than the Gamma test result led us to believe but very much in the same ball park The results of the test with k 2 20 are shown in Figure 3 5 Here we can see that the agreement between the predicted blue trace and the actual test data green trace is very close The red trace indicates the error
39. e particular output The goal of model identification for a particular output is to choose a selection of inputs that minimises the asymptotic value of the modulus of the Gamma statistic All things being equal this should result in a model which has minimal MSError when used to predict the output using input data not seen in the model construction process What happens if the final conclusion is that the noise variance on the output we are trying to predict is unsatisfactory We can attempt one or all of the following Increase the accuracy of measurements of both the inputs and the outputs The effective noise variance on the output may be the result of measurement error on the inputs Ask if we have included all the principal causative input variables liable to affect the output If some obviously important factor has been missed then this may well explain why we are currently unsuccessful in predicting the output variable For a time series prediction we could increase the rate of sampling or consider if there are other time series which may have predictive value for the time series we are interested in predicting such time series are often called leading indicators One reason the Gamma test is so useful is that it can immediately tell us directly from the data whether or not we have sufficient data to form a smooth non linear model and how good that model is liable to be If the result is that the error of prediction is too h
40. e searched over the previous 15 years O Results Visualiser Select output Output 1 x Frequency Custom Chart XSeries Unique Data Pc v Overlay Y Series EN 20 40 6 80 100 120 140 160 180 200 220 240 20 y Unique Data Points Garma 0 0050 010 0150 020 0250 030 0350 04 Figure 2 35 The M test graph for the sunspot Figure 2 36 The frequency histogram of all data data using the best embedding of length embeddings of length 15 using the sunspot 15 data The best embedding found was 001001000010111 Here the most recent data comes last So this embedding says that to predict this year s sunspot activity x t we should use the data x t 1 x t 2 x t 3 x t 5 x t 10 and x t 13 an embedding of dimension six It is interesting to note the bimodal distribution of the Full Embedding Histogram of Figure 2 36 The bimodal distribution is partly explained by the observation that only 2 38 of the embeddings with a Gamma statistic gt 0 008 include x t 1 as compared with 99 8 of those having a Gamma statistic lt 0 008 Put plainly this says that the most important predictive factor for the sunspot activity this year is the value for last year It is also interesting to see which variables appear in the best few embeddings ERE These indicate that the last few years plus messe EEE the value approximately one 11 year cycle discs bez ei back plus a value about half way through the Model Test
41. ear data The Increasing Near Neighbours plot for pmax 3 to 30 is given in Figure 2 13 This suggests the best estimate for the Gamma statistic is obtained at around pmax 17 The M test result of Figure 2 14 was obtained starting at M 50 and increasing M to 500 in steps of 10 This consistently gives a Gamma statistic of around 0 07 but ideally as the graph has not yet settled to an asymptote we should need more points to obtain an accurate estimate for this noisy 1 dimensional data The scatter plot in Figure 2 15 contains points with small 6 but large y which also supports the conclusion At the same time the regression line fit is rather poor The 3D histogram in Figure 2 16 shows partial indicators of an empty wedge and supports the general conclusions that the data is noisy The same is true ofthe angle histogram in Figure 2 17 Finally the Moving Window Gamma test using a window size of 300 in steps of 10 in Figure 2 18 consistently shows a Gamma statistic between 0 072 and 0 076 These results together indicate that we have noisy non linear data but that model construction is quite feasible The winGamma User Guide PERFORMING AN ANALYSIS Version 18 Jan 2002 Results Visualiser B Results Visualiser Select output Output 1 zl Select output Output 1 zl Custom Chart X Series Near Neighbour zl x Near Neighbours v Gamma and Standard Error Unique Data Pe zl Unique Data Points v Gamma lt Primary Y Series H E
42. ece eee eee 15 Figure 1 7 The Analysis and Data Set Manager windows after performing the initial judi MR rc ET 18 Figure 1 8 The noisy sme data 2r ic upvu ga 3 be ee ara EEG GU EA Edu 19 Figure 1 9 An M test on the noisy sine data 0 eee eee eee eee 19 Figure 1 10 The first 100 points of the Hen500 asc time series llle 20 Figure 1 11 The surface which defines x in the Henon map as a function of the two previous SATB Ls s e Pueden dC eon A 809 ADR d PE QURE P RO d n o oca pu ul og A cara E 20 Figure 1 12 The distribution of points in the input space for the Henon map 20 Figure 1 13 The result of an Increasing Embedding on Hen500 asc with a maximum of 10 inputs using 10 nearest neighbours 2 0 0 cee eee e 21 Figure 1 14 The result of an M test on Hen500 asc with 2 inputs using 10 nearest neighbours sched G Mabe Lo E eg os ONES al oue aa dash EORR pea Sach lh atta a da 21 Figure 1 15 The scatter plot for 280 test points on Hen500 asc with 2 inputs using 10 nearest Berg iDOUF S sews nas gi eee I CP EROS Ce ger der UD SURE Oe Ree ERS Cea Tee 22 Figure 2 1 Main Features of the scatter plot and regression line 28 Figure 2 2 Modulated sine curve used to generate the Input Output file ModSin5000 asc 29 Figure 2 3 Scatter plot with pmax 100 for ModSin5000 asc 0 0 eee eee 29 Figure 2 4 The variation of Gamma and SE as the number of near neighbours increases 30 Figure 2 5 M tes
43. election of inputs because real measurements are subject to error and there is no point in building more and more accurate models for example by using the noise cancelling features of local linear regression because the predictions of the model will never agree with our measured data unless the measurement error is decreased for example An exception to this might be if are trying to get some idea about an underlying theoretical model and winGamma can help in this respect but determining a theoretical model as opposed to an accurate numerical model lies outside the competence of winGamma How should I choose between a local linear regression LLR method and a neural net method of model building Nets take a long time to train but may generalise better than LLR in regions of the input space where data is sparse A high Gamma statistic on the training data may make neural network training even more difficult If data is densely distributed over the input space then LLR may be a better choice in this situation The particular application also has an influence on which may be the best modelling tool to select For example to learn new data it may be necessary to retrain a neural network from scratch which is time consuming whereas dynamic LLR can easily accommodate new training data Local linear regression models are very fast to build but take relatively longer to query because a kd tree is used to find the near neighbours of the query poi
44. ending up with a long GA run 6 If a better embedding is found then repeat steps 2 3 and 4 to refine those conclusions The winGamma User Guide PERFORMING AN ANALYSIS Version 18 Jan 2002 Time series files Time series analysis is complicated by the fact tat we probably do not know how far back in time we should look to build our prediction model This initial decision is not irrevokable and should be guided by some degree of commonsense analysis on what is likely to be the case for the given data set and how much data is available E g For a single time series with annual periodicity where the samples are weekly we might set the number of inputs to 104 the equivalent of two years However 104 dimensional space has a lot of room up there and we should need a data set going back many years to make this worthwhile If only a few years data is available then perhaps we should first consider a model over the last several months or weeks 1 Load the data and do not initially normalise if it is a single time series Set the number of inputs to a reasonable maximum in the light of the data and the number of outputs initially to one Now perform a simple Gamma test on the full data set if not exceedingly large with the default number of near neighbours set to 10 to get an initial idea If the data set is very large use a subset of the data for initial experiments 2 Run an Increasing Embedding test to determine a likely embedding dimen
45. ents are Pentium processor 133 MHz or preferably faster RAM 32 64 Mbytes The amount of memory you will need to run winGamma is not really constrained by the program so much as the size of the data sets that you wish to analyse With the possible exception of the neural network training algorithms the theoretical average case computation times of the main algorithms in winGamma scale like O MlogM where M is the number of rows in the data file However under some conditions some algorithms in winGamma may require quite a lot of memory to achieve the theoretical scaling An example is Increasing Near Neighbours when pmax is large Suppose we consider solar csv sith 10578 rows of three numbers each and set pmax 100 This demand will require approximately 0 25 Mgbytes for the data 0 25 Mgbytes for the kd tree but more than 4 Mgbytes for the 10 numbers which constitute the list of 100 nearest neighbours for each of the 10578 input vectors in the data file To perform a Gamma test each of the near neighbour indices must be instantly available and they could be anywhere in the range 1 10578 If the system has less than 4Mgbytes of available RAM then it will have to keep paging data in and out from the hard disk This will dramatically slow the algorithm and may in fact render the entire computation infeasible If you observe a large amount of continuous paging disk activity then a Close down all other applications b Consider if it is feasible to
46. er previous cycle give the best results This is SEES rather impressive since the software has no way of knowing about sunspot cycles If we run the Gamma test on the six inputs one output I O data file constructed using this mask we get Gamma 0 0015 and V xo 0 036 SE 0 00093 Note the M test did Era of Figure 2 35 indicates that there is not 208 210212 214 218 218 220 222 224 226 228 230 232 234 236 238 240 242 244 246 248 250 252 254 256 258 260 262 264 266 really enough data the graph has not e stabilized Therefore if we construct a model Figure 2 37 A test of the LLR model on the data and test on unseen data we might expect to set SunPairs asc blue predicted green actual red get a higher MSError than the estimated RUD n a 47 The winGamma User Guide PERFORMING AN ANALYSIS Version 18 Jan 2002 gamma value If we now predict the last 59 years data using local linear regression with k 60 near neighbours and a threshold of 0 0001 we shall see how to do this in the next chapter on the basis of all the previous years we obtain Figure 2 37 which gives a MSError around 0 007 In cases such as this where there is insufficient data it is not uncommon to see a MSError on unseen data around an order of magnitude greater than the Gamma statistic 2 14 Handling multiple time series In later releases we plan to include a TimeSeries Editor to facilitate the direct manipulation of multiple time s
47. eries However using a combination of winGamma and Excel at present it is possible to accomplish multiple time series manipulation relatively easily Suppose we have several time series TS1 TSm which we wish to use to predict a target time series Target Step 1 Suppose the time series are in an Excel file which is structured as follows Table 2 4 Excel file for multiple time series Date TSr TS2 TSm Target 10 07 1981 10 39 132 06 606 8 10 14 1981 15 5 10 34 132 92 606 8 10 21 1981 155 1038 130 43 626 a o e e de e In Excel delete the date column because this is not numerical that can be used as input for winGamma notice now that the only factor which preserves the time relationship is the order of the rows Hint It is important when dealing with multiple time series that all the data in a row is sampled at the same time If one measurement is sampled weekly and another monthly then we can use linear interpolation to construct weekly data samples for the monthly sampled data Step 2 Save the file from Excel in CSV format you should include the first row of text descriptors of the time series as winGamma can handle these and they will be useful later Step 3 Load the file just saved into a text editor which can SaveAs ASC DOS text files and search and replace the commas separating the numeric data only Do not the replace the commas separating the commas separating the text he
48. erve but is not actually the best way to use dynamic LLR It is better to start with a reasonable training set size because then the initial kd tree a data structure used extensively by winGamma will be more balanced and query times will be reduced 33 The winGamma User Guide BUILDING AND TESTING A MODEL Version 18 Jan 2002 Under the circumstance it is not surprising that if the same test data is presented to the model a second time the MSError will reduce dramatically 3 3 Two layer back propagation Modelling E ditor Modelling E ditor Model type Mw propagation Two Layer Neuraiad Model type Backpropagation Two Layer Neura Network Parameters rraring Parameters Network Parameters Traring Parameters gt Network min Ri Training Parameters 4 Number of nodes in first layer Initial leaming rate E F fo 25 Network initialisation time 5 Number of nodes in second layer Momentum 59s eS j 5 a o5 p Regularisation Train to 3 0 001 pam Target MSE 3m P fo 00105462113952337 Aecalc Network training lime inutes Figure 3 1 The first set up menu for two Figure 3 2 The second set up menu for two layer backpropagation layer backpropagation This option uses the standard backpropagation algorithm to produce a two layer feedforward neural network With all the neural network training algorithms one should note the option to recalculate the
49. ese experiments and the interpretation of their results fully in Chapter II For the present we shall simply illustrate the basic Gamma test experiment The Data Set Manager shows the data that has been loaded as in Figure 1 5 where the windows have been tiled Because data files may be very large the data rows are divided into pages of 100 rows each In Figure 1 5 the first page has been selected Each column represents a column of inputs or outputs and is labelled as such The first four rows give the Mean Standard Deviation Minimum and Maximum of each column for the whole of the data selected for analysis The name of the current data file is also displayed at the top of this window Handy Tip Note that most of the windows and sub windows including the column separators in the Data Set Manager data display can be resized using click and drag To perform a Gamma test select the Analysis Manager and then Experiments Highlight Gamma test and select New We can now toggle between the Experiment tab and the Mask tab The only option to be set from the Experiment tab in this experiment is the number of near neighbours For 13 The winGamma User Guide GETTING STARTED Version 18 Jan 2002 the present leave this set to 10 The Mask tab is used to select which inputs to include Leave this set to 11 i e both inputs are included winGamma File Edit Transform Options View Window Help C o m m I Analysis Manager New Del
50. ession of models based on 1 Input the previous value 2 Inputs the previous two values etc up to the maximum number of inputs we have set which is 10 where in each case the output is the current value Results This gives us a succession of Gamma values which we can graph by clicking on Graph The result is shown in Figure 1 13 Here it becomes clear that the best of these models i e the one having a Gamma statistic closest to zero is the one which uses just the two previous values The Gamma statistic for this model is approximately 0 000161 which is very close to zero The Vratio is 0 0001648 which again is close to zero 7 Now that we have identified the relevant inputs pull down the Transform menu and click on Transform the data set Under the Time Series Options select 2 inputs and 1 output and then leave the proportion of data set for analysis set to 1 498 8 Next in the Analysis Manager under Training Set Analysis select M test and then click on New In the Experiment Editor click on the M test tab and set the Initial sample size to 10 leave the final sample size set to 498 and set the step size to 10 Now click on Execute We should like to see how stable the Gamma statistic is and how much data we are likely to need to get a good quality model Finally when the results window comes up click on Graph to see the result of the experiment This is shown in Figure 1 14 Results We see from the graph that we coul
51. ete Afalvse Graph Madel Test Huer vw hat Predict Experiments Models E e Training Set Analysis IEEE RC Increasing ne sneighbours e M Test Moving window gamma test H e Model Identification Data Set Manager Page p File solar csv 6 5962 2 6583 2 2676 5 2233 0 38343 5 928 15 679 28 22 CD CO 7 CD Cl e CO n2 Figure 1 5 The Analysis and Data Set Managers after loading a data file When these steps are complete click on Execute Under the Analysis Manager the Settings window will now show the settings for the current experiment This is shortly augmented by a Results window which shows the results of this experiment We can switch between the Settings and Results windows using the appropriate tabs These results for the single output are presented in a Results Settings window along a single row because there is only one output and are shown here in the first column of Table 1 1 If there is more than one output the software generates a similar set of results for each output Finally Transform the data and repeat the experiment to obtain the scaled results in the last column of Table 1 1 1 3 1 Interpreting the results To interpret these results it helps to have some idea of how the Gamma statistic is calculated We shall describe this more fully in Chapter II but for now it is enough to know that the Gamma statistic is calculated by determining a regression line based on near neighbour statist
52. ewer 1 s and thereby place more emphasis on simpler models Note the three weightings selected for GA fitness should sum to 1 Run Time 5 minutes The approximate maximum time selected to perform the GA Setting the population larger may improve the final fitness of the best mask found but only if a large run time is permitted For long masks i e a large number of inputs and large data sets the GA will require runs of several hours 2 7 3 Hill Climbing In hill climbing a mask is taken default is all ones for the current number of inputs and each bit is flipped in turn calculating the Gamma until the end of mask is reached This is repeated until no single it flip gives an improvement on the Gamma This is a relatively fast heuristic but takes longer than a sequential embedding 2 7 4 Sequential Embedding 33 The winGamma User Guide PERFORMING AN ANALYSIS Version 18 Jan 2002 Here a single pass through the current mask is made flipping each bit only if there is an improvement over the Gamma statistic obtained with the original mask Again default is a mask of all ones equal to the length of the current input vector though as in the previous method an initial mask of any kind can be used provided its length does not exceed the current number of inputs This is very fast 2 7 5 Increasing Embedding The Increasing Embedding algorithm starts with the mask obtained by taking only the rightmost input in the case of a time series
53. files of data for testing When the analysis data set doesnt include the test data or further tests need to be made on a model When should I use the moving average option Usually when you have plenty of data and want to determine if the number of data samples used to estimate the Gamma statistic gives a stable value over a range of different sample sets of the same size This test is also useful to investigate if the underlying dynamics is itself varying When should I use the differential option It may improve the MSError for difficult time series Which input is the differential or moving average input When these options are activated the new data column is placed in the highest numbered positions with differential first and moving average last This can be confirmed by placing the cursor over the vertical column and dragging it wider thus revealing the applicable legend 76 INDEX Analysis Manager window 13 angle histogram 29 Data 9 data model 9 Data Set Manager 13 embedding 32 embedding dimension 21 Embedding Histogram 32 Excel Macro 62 Experiments window 13 full embedding 25 32 GA Fitness 71 Gamma statistic 15 Gradient 16 heuristic search techniques 25 input columns 9 mask 32 Normalization standard normalization 50 observations 9 output columns 9 over training 9 71 Over training definition 71 Project definition 50 Query 58 62 Results window 14 Standard Error 16 Vratio 16 Whatlf 58
54. function generating the data provided this is itself a smooth function You can easily see this by using the winGamma Whatlf facility on LLR models built from increasing sized data sets for example on the data from Sin500 asc LLR can produce very accurate predictions in regions of high data density in input space but it is liable to produce unreliable results for non linear functions in regions of low data density In other words LLR does not generalise well but is a very good interpolative tool if we have large amounts of data There are three user settable parameters in LLR the number of near neighbours whether or not to include a constant term in the linear model and a threshold value for filtering the local eigenvectors The choice of the number of near neighbours k in LLR is quite critical If the noise level on the output i e the asymptotic Gamma statistic is low then some small multiple of the number of inputs should suffice If the noise level on the output is high then k needs to be larger to obtain better noise cancellation Unfortunately if the unknown function f we seek to model is highly non linear has regions of high curvature then unless we have a very large amount of data setting k large may mean that in the region of these k points the assumption that the unknown function can be locally approximated by a linear model may be false In this case the resulting predictions would be inaccurate We have not yet developed rules of
55. grees Irradiance is 10 Similarly we can ask 58 The winGamma User Guide BUILDING AND TESTING A MODEL Version 18 Jan 2002 How does the power output vary when the Irradiance is fixed at 10 and the Temperature from 1 to 15 The answer is given in Figure 3 7 Here the result is somewhat different One interesting feature is the slight rolloff of power output with increasing temperature This is a real effect and is a consequence of the physics of solar cells 3 6 1 2 A histogram of prediction errors for LLR model If we save the data produced by the results of the LLR test then we can examine an error histogram for the predictions This is shown in Figure 3 8 The vertical gridlines are one standard deviation either side of the mean which is close to zero This is the final test of our model 3 6 2 Building and testing a neural model Figure 3 8 An error histogram for the LLR test We can now repeat the model building process using a neural model 3 6 3 Visualising the data For a 2 Input l Output data set we can visualise the model as a 2 dimensional surface and using suitable software plot this surface directly from the data Of course in higher dimensional spaces such graphical realisations are not possible Moreover if the data is very noisy such a surface will be very jagged and not much use as a model Nevertheless now that we have finished studying solar csv it would be interesting to see what
56. he structure of the data as in multiple Input Output data sets the only parameter to optimise is the number of near neighbours often denoted by pmax It is a remarkable fact that for many data sets the default of pmax 10 nearest neighbours is often nearly optimal A suitable size for pmax in the Gamma test principally depends on two factors The number of data samples M if M is large the local number of data points close to a given point can be expected to be high The ocal curvature of the surface described by the unknown function f other things being equal for a surface with high curvature we cannot afford to take neighbours too far away so that pmax will require to be smaller Systematic ways to determine the best choice for the number of near neighbours are described later in section 2 4 Note that the size of pmax in modelling the unknown function f using local linear regression is determined by other factors described in section 3 2 Whilst for the Gamma Test it is usually the case that we want to take pmax small for local linear regression at high noise levels we will need to take pmax much larger 2 3 The Gamma Test analysis graphs After performing an experiment highlight the row containing the Gamma result to be scrutinised and click Analyse Clicking on the tabs will provide the other plots that are discussed below In an experiment where there are multiple Gamma results the graphs and plots will relate to the highligh
57. his assessment Now click on Analyse This enables us to see three analytical graphical displays which are described more fully in Chapter II The first of these displays is the Gamma scatter plot and regression line of Figure 1 6 The other two tabs give a 3D Histogram and an Angle histogram These are different ways of viewing the data in the scatter plot Results Visualiser How stable is the Gamma statistic with 10 near em ke neighbours as the number of data points varies We can answer this question by clicking on the Experiments tab and then highlighting M test This NK will run the Gamma test for an increasing number _ oveisyseies M of data points Now click on New to begin Z setting up the M test leave the number of near neighbours set to 10 and click on the M test tab Set the initial sample size to 10 the final sample size to 500 and the steps size to 10 Now click on Unique Data Pc Y Unique Data Points v Gamma 50 100 150 200 250 300 350 400 h 450 500 Execute to begin the Experiment After the Results Dige pan Pon window comes up click on Graph to obtain a graph of the Gamma statistic values against the Figure 1 9 An M test on the noisy sine data 19 The winGamma User Guide GETTING STARTED Version 18 Jan 2002 number of data points This is shown in Figure 1 9 and we can see that after around 425 points the graph is fairly stable The fact that the data is rather noisy
58. hniques Local linear regression Dynamic local linear regression Neural networks using various types of learning algorithm Local linear regression models are fast to construct and quite fast to execute a query Local linear regression models can also be easily updated as new training data becomes available which is not the case with neural networks where a prolonged extra period of training or starting training all over again may be required to modify the model on the basis of new data Indeed winGamma also offers a dynamic local linear regression option which is exactly local linear regression with dynamic updating this option is quite useful for time series prediction Usually local linear regression is extremely accurate in parts ofthe input space where the training data density is high However local linear regression will not generalise well to parts of the input space for which training data is sparse Neural network models take time to construct but in parts of the input space where data is sparse tend to generalise better than local linear regression It is often quite hard to get a neural network to train down to a very small Gamma statistic say 10 or 10 which can easily happen with zero noise dynamical system time series i e it may take several attempts each of which takes a long time However neural networks can make predictions at blinding speeds compared with local linear regression based algorithms so for some applicatio
59. hs ee Qux MESE X REAR eR SEX WE OE ea Y SEMEN FA SEE 67 Prediction file data 22 252 4 Aig beess i ps ERGO E REESE AERE 68 APPENDIX III Using the Mathematica 4 0 files 0 cece eee eee 69 DataGemn 4 cen ae he v e Ee Sete Sean etr ux ERAS dS 69 DataAnal mb 4 2 5 46b 2226 eh asee sie BUS toy heces EPS E EE hares rid e 69 mathlinkGamma nb 0 0 cece eee cece e ees 69 NetReader MD mte tt DER EN NES o NN RE NAE 69 APPENDIX IV Generating test files uuovak ead dep IO nee aih od PEN eye 70 Generating your own data files svo ua LAE RR ood on EAE REALES 70 Creating data files using Excel esie ure ev nb yee wee oe eae 70 APPENDIX V DefnnJtiols eve RE eI ape yeah EE eleeqam eK eie RE edad 71 APPENDIX VI Frequently asked questions 0 0 cee eee eee eese 13 INDEX bI A Qa Det pie FLA ED EE R QC DRE O eb b d 71 List of Figures Figure 1 1 Toggle inputs to outputs as required when loading a csv file as Input Output data bord iS Anon adi EA Ke SERA OAS EA ovedte dx odutib diede dre dS Ud M aca 11 Figure 1 2 Selecting the number of inputs and outputs per time series 11 Figure 1 3 The Normalise check Box 45 023 vxo ete us cigs Maes vx da acsi eee oe 12 Figure 1 4 Selecting a proportion of the data for initial analysis 15 Figure 1 5 The Analysis and Data Set Managers after loading a datafile 14 Figure 1 6 The Gamma statistic and the Gradient Slope c
60. ics derived from the data see Figure 1 6 14 The winGamma User Guide GETTING STARTED Version 18 Jan 2002 Gamma The first row of Table 1 1 gives Table 1 1 Gamma test results with pmax 10 for unscaled and scaled the Gamma statistic pmax 10 solar csv data for the output as evaluated over uL d e analysis in h Unsealed Scaled this case the whole data set As onecansec from Fieuws 106 tic Gamma statistic is actually the vertical intercept of the 0 000760 0 000785 regression line in the figure This is the estimated variance of the errors for any smooth model built on the data Since the output variable range is approximately 0 30 this is a relatively small error variance It oe on this data will have a standard 7 deviation of the prediction error of about 0 020761 0 144 on the unscaled data which is about 0 5 of the range In general it is helpful to distinguish two cases First where the true noise variance is zero In this case the asymptotic Gamma statistic should approach zero and there is no limit to how good a model we can build provided only that we have more and more data of arbitrarily high precision For example this can happen with artificially generated data for chaotic time series Second and more realistically where the true noise variance is positive In this Figure 1 6 The Gamma statistic and th case the asymptotic Gamma statistic yq dient Slope should a
61. igh no matter how much data we are given then we must address the above issues For each choice of inputs investigated as the number of data points increases we attempt to establish the asymptotic Gamma statistic for each output We then choose the set of inputs for a particular output that has the minimum asymptotic Gamma statistic this is known as model identification Having established the best selection of inputs for each output using the winGamma software models may be built by e Static local linear regression fixed model Dynamic local linear regression model updated as new data becomes available or by using one of four different types of neural network training algorithms d Convergence in probability Because of sampling error if the variance of the noise level on an output is very small the Gamma statistic may sometimes be negative even though a variance can never be negative If this occurs we use the absolute value or modulus of the Gamma statistic 10 The winGamma User Guide GETTING STARTED Version 18 Jan 2002 Two layer back propagation Meta backpropagation Not included in the Beta release Conjugate gradient descent BFGS neural network Predictions on new input data for which the outputs are unknown can also be made using one or more of the models 1 1 2 The range of applicability The software is designed to analyse data with the goal of producing a near optimal smooth function fr
62. iginal data set which have Ix 1 x j small i e x 1 and x j are close together but their corresponding output values y have ly 1 y j large i e y i and y j are far apart This is very bad from the viewpoint of constructing a smooth model It may be a reflection of a high intrinsic noise level on y a high gamma or it may just be that there is no smooth underlying model Gamma Scatter Plot 200 000 400 000 600 000 800 000 1 000 000 1 200 000 delta Figure 2 1 Main Features of the scatter plot and regression line An example where the underlying model is not intrinsically smooth might be a logic function of the input variables e g XOR or m bit parity In m bit parity the inputs are the vertices of the m dimensional unit hypercube and the outputs are 1 or 0 In fact one can put a smooth surface through these points but this is a rather meaningless exercise Problems with a large number of discrete input or output variables are best tackled via a decision tree approach rather than trying to use smooth modelling techniques The scatter plot can also give important clues on the nature of the data For example it can happen in some control applications that the system being modelled goes through two or more different dynamical regimes In one instance the scatter plot revealed that there were really two different regression lines each corresponding to a different dynamical regime Moreover each regime corresponded to a distinct p
63. iple Time Series winGamma assumes that non determinism in a smooth model from inputs to outputs is due to the presence of statistical noise on the outputs Not all phenomena that one might seek to model fall into this category For example if the outcome that one is trying to predict from observations is highly probabilistic then the model produced by winGamma will not be satisfactory as a prediction tool e However the software is able to detect this situation The models that winGamma is designed to produce are of phenomena more exactly outputs that are smoothly determined by the input variables Mostly the limiting factor on the predictive accuracy of the model will be measurement noise or insufficient data For a given data set the winGamma software executes the Gamma Test which estimates the variance of the noise on each output This will be an estimate of the best MSError that a smooth model can achieve for the corresponding output Inputs and outputs should be continuous variables See Appendix V for definitions Tt will be reflected in a high Gamma statistic or a Vratio close to 1 The winGamma User Guide GETTING STARTED Version 18 Jan 2002 The estimate of that part of the variance of an output that cannot be accounted for by a smooth data model is called the Gamma statistic As the number of data samples increases the Gamma statistic invariably approaches an asymptotic value which is the variance of the noise on th
64. istributions of data like this can be very helpful in high 20 The winGamma User Guide GETTING STARTED Version 18 Jan 2002 dimensional input spaces as it often means that we need less data to build a good model than would be the case if the data were uniformly distributed over the whole space It is precisely the surface of Figure 1 11 which is the model that we can seek to construct using winGamma We could take the time series and create a 2 Input 1 Output data structure x x gt XQ In fact any time series that evolves according to some smooth iterative or dynamic process can be treated this way provided only that we can determine the number of previous values of the time series required to predict the next value this is called the embedding dimension In this example we shall pretend that we do not know the embedding dimension and show how winGamma can be used to get some idea of which previous inputs are likely to produce a good model Note that the data in the file Hen500 asc is high precision and not subject to noise 1 5 2 1 The basic steps 1 Load Hen500 asc with Open Analysis Data Set 2 Set the number of inputs to 10 in the Time Series tab 3 Do not enable Normalisation in the check box Since the data is a single time series and each sample is comparable we should not expect much gain from scaling 4 When prompted to select a proportion of the data set for analysis use all the data 1 490 for
65. izontal With pure random non smooth data the slope of the regression line will gradually increase as the number of data points M is increased this is because the continuity condition is not satisfied 34 The winGamma User Guide Taken together particularly with Vratio so close to one these are clear indicators that it is pointless to try to model the data with a smooth function Next we examine the standard analysis tests in turn The Increasing Near Neighbours plot for pmax 3 to 30 is given in Figure 2 7 This suggests the best estimate for the Gamma statistic is obtained at around pmax 10 The M test result of Figure 2 8 was obtained starting at M 50 and increasing M to 500 in steps of 10 This consistently gives a Gamma statistic of around 0 3 but ideally PERFORMING AN ANALYSIS Version 18 Jan 2002 Table 2 1 The results of a simple Gamma test on the file Ran500 asc for unscaled and scaled data po nscale Scaled Start 1111 1111 as the graph has not yet settled to an asymptote we should need more points to obtain an accurate estimate for this 4 dimensional data The scatter plot in Figure 2 9 contains points with small 6 but large y which also supports the conclusion At the same time the regression line fit is rather poor The 3D histogram in Figure 2 10 shows no real indicators of an empty wedge and supports the general conclusions that the data is extremely noisy The same is true of
66. le choice is to start at 100 in steps of 100 until the end of the data Results The M test graphs of the Gamma statistic together with the Gradient are shown in Figure 2 29 From this we see a good asymptote and conclude that with 8 inputs a good model can be obtained using around 3000 points It also looks likely that we have an essentially zero noise time series 43 Results Visualiser Select output Output 1 v Unique Data Pc Primary Y Series Gamma z Overlay Y Series E G Unique Data Points v Gamma and Gradient 500 1000 1500 2000 2500 3 00 3500 4000 4500 Unique Data Points h Gamma Gradient Figure 2 29 The M test graph pmax 10 number of inputs 8 for the delayed Henon map The winGamma User Guide PERFORMING AN ANALYSIS Version 18 Jan 2002 5 Can we get a better Gamma statistic by discarding still more of the inputs To answer this question we run a Full Embedding on 7 inputs To do this we have to transform the data again Results If we make a small table of the best 5 masks found we obtain Gamma 2 3228E 6 1 3276E 5 1 888E 5 2 176E 5 2 438E 5 Mask 0001101 1101100 1111111 0001100 1011101 From this we infer that lags 3 and 4 remember we have to count from the right are very important but that the marginally best model should be obtained using lags 1 3 and 4 Results Visualiser It is worth examining the Embedding Select output Output 1 v Histogram associated wi
67. lowed by one or more spaces The end of the input vector is signified by a comma What follows the comma is one or more outputs separated by one or more spaces The last number on a line should be followed by a carriage return linefeed There must be the same number of data fields before and after the comma on each row Prediction file data 68 APPENDIX III Using the Mathematica 4 0 files A number of Mathematica files for data generation manipulation data analysis and model testing are supplied with winGamma There is also a C code executable MathLink file which can be used to execute the Gamma test from within Mathematica To use these files you will need to have Mathematica installed and be familiar with Mathematica notebooks At a later stage it is hoped to supply equivalent files in Matlab DataGen nb Data Generator This file enables the creation of Input Output files and Time Series files of data with our without added noise In includes a large number of examples and shows how every test file used in this manual was created DataAnal nb Data Analyser This file is useful for producing graphics such as histograms and performing various types of supplementary data analysis GammaTestProject exe This is a C code executable which communicates with Mathematica via MathLink and enables a variety of Gamma test computations to be called directly from Mathematica It cannot be executed as a standalone program mathlinkGamma n
68. lso be positive and there will come a point where using more data to build our model will not actually improve the quality of the predictions when compared with the measured values of the output Delta In the case of a positive asymptotic Gamma statistic we can determine the minimum amount of training data required to build a smooth model with this MSError using the M test described in section 2 5 15 The winGamma User Guide GETTING STARTED Version 18 Jan 2002 Gradient The Slope or Gradient is the slope of the regression line in Figure 1 6 used to calculate the Gamma statistic It is actually a rough measure of the complexity of the smooth function we are seeking to construct In this case the gradient of A 0 244 indicates that the output is a rather simple function of the two inputs It is generally best to look at the Gradient for the scaled data since this refers to a standardized output range Like the Gamma statistic the Gradient will eventually asymptote to a fixed value However the number of data samples required to get a stable asymptote for the Gradient will usually be much larger than the number required to get a stable asymptote for the Gamma statistic Standard Error SE This is the usual goodness of fit applied to the regression line in Figure 1 6 If this number is close to zero we have more confidence in the value of the Gamma statistic as an estimate for the noise variance on the given output In this case an
69. means we should try to optimise the number of near neighbours for the Gamma test if we wish to obtain a more accurate Gamma statistic and we shall see how to do this in Chapter II 1 5 2 A chaotic Time Series 3A EIER 1 5 Figure 1 10 The first 100 points of the Figure 1 11 The surface which defines x Hen500 asc time series in the Henon map as a function of the two previous values Here we use the file Hen500 asc This file contains time series data generated by iterating the Henon map It is described in more detail in The Gamma test and how to use it a practitioners guide To get some idea of what the time series data looks like we graph the first 100 points of the time series using any convenient software as in Figure 1 10 Although this time series looks quite unpredictable nevertheless the underlying model which takes us from two successive values to the next is a smooth function of the two successive inputs and therefore does not violate the requirement of the Gamma Test see Figure 1 11 Figure 1 12 The distribution of points in the input space for the Henon map A very important factor to consider when building a non linear model is the distribution of sample points in the input space In some cases these points will be uniformly distributed but in many cases this will not be the situation If we plot the distribution of the points x Xn for the Henon map data from the file we obtain Figure 1 12 Peculiar d
70. n 18 Jan 2002 In some data sets the same input vector may occur several or many times This indicates how many distinct input vectors are present in the data see the discussion on zeroth near neighbours below Evaluated Output Indicates to which output the results relate In a file with multiple outputs all these results are calculated for each output Zero th Near Neighbours In some data sets the same input vector may occur several or many times If an input vector appears multiple times then if it has the same output value s it might be construed as a repetition or it may be a separate independent observation In the first case there is no extra information and the data vector should be deleted In the second case there is useful information in the two vectors because they are telling us that for these inputs the outputs are identical and so presumably subject to low or Zero noise variance If one or more outputs are different for the same input vector then again there is useful information because enough vectors of this type could give us an immediate grip on the noise variance Therefore because it is important for an analyst to know if the same input vector occurs multiple times winGamma provides this information by stating the maximum number of non unique input vectors If this number is small in relation to the size of the data set it can safely be ignored on a first pass If it is large then the data should be subjected to some a
71. n this experiment was FTP ed from ftp address ftp santafe edu directory pub Time Series data Its origin normalization and training test regions are described in Weigend 1990 The data consists of 280 points representing sunspot activity over the period 1700 1979 and was used in Weigend 1991 The range of the data has been scaled to 0 1 and we found the variance to be 0 0410558 Figure 2 34 shows the variation of sunspot activity over the full range of the data Itis known that the primary sunspot cycle is approximately periodic over 11 years Other shorter and longer cycles are also known For radio propagation the short period cycle of 28 days is particularly significant The data used here is collected from telescopic observations projected onto a white paper card The sunspots are counted and classified by size and a correction factor applied depending on the magnification of the telescope The virtue of this data is that it has been regularly collected since 1700 Of course if one were really interested in predicting sunspot activity much more accurate data is available The data provided is often used as a test of prediction techniques and can give a reasonable model of gross sunspot activity 46 The winGamma User Guide PERFORMING AN ANALYSIS Version 18 Jan 2002 Selecting a best embedding If we are prepared for a several day run we can use the Full Embedding option of the software to search for a good embedding In this example w
72. nalysis outside of winGamma Upper 95 Confidence Lower 95 Confidence In the case where zeroth near neighbours are present these results are the lower and upper bounds at the 95 confidence levels for the Gamma statistic estimated directly from the zeroth near neighbours Unless the data file has many repeated input rows these values can be ignored If the file has many repeated inputs then these values can be compared with the normal Gamma statistic which is computed in an entirely different way 1 4 The basic controls of winGamma The use of these options will discussed fully in Chapter II The Analysis Manager Experiments These are options used to determine the Gamma statistic and to investigate how reliable this statistic is i e to determine the quality of a model which might be built using the data and a given selection of inputs To invoke any of these options after loading a data set simply select the Analysis Manager and highlight the option required Then click on New For any particular option there are probably other parameters which require to be set before invoking Execute Gamma test Finds the Gamma statistic and other relevant measures 17 The winGamma User Guide GETTING STARTED Version 18 Jan 2002 Increasing near neighbours Finds how the Gamma statistic varies with the number of near neighbours used to compute it M test Shows how the Gamma statistic estimate varies as more data is used to compute it
73. nces of the error surface in weight space which in most cases gives faster convergence at the expensive of a more complicated algorithm and more memory What do all the fields associated with a Gamma Result mean See section 1 3 1 of this manual What does a high gradient suggest If there is enough data to give a stable Gradient asymptote then a high Gradient computed values on artificial test sets can come out as high as 20 000 suggests a complicated unknown function with on average regions of high curvature Why is the Vratio useful It provides a standardised estimate of the noise which is independent of the output variable range What is the use of the Standard Error Ittells us how reliable the Gamma statistic is as an estimate of the variance of the noise on the output What file formats are permitted for data to be analysed by winGamma See Appendix II How much data should I use for training If the Gamma statistic is asymptoting to zero you can use as much data as is practical and models with MSErrors of order 10 are quite feasible 75 The winGamma User Guide APPENDIX VI Frequently asked questions Version 18 January 2002 If the Gamma statistic is asymptoting to a positive value a good rule of thumb is to use as much data as will give a standard deviation the square root of the variance of the Gamma values about the asymptote of around 10 of the asymptote value on the last 10 of the data When should I use external
74. ns it is well worth the time and effort to construct a neural model 3 2 Local linear regression To make a prediction for a given query point in input space local linear regression LLR first finds the k nearest neighbours of the query point from the given data set where the number k is supplied by the user and then builds a linear model using these k data points Finally the model is applied to the query point thus producing a predicted output Because of the way winGamma analyses the data to compute the Gamma statistic the k nearest neighbours of any point in input space can be found very rapidly Consequently local linear regression using the k nearest neighbours in the training data of the query point can be accomplished quickly Thus local linear regression is a very fast and capable predictive tool LLR is most effective in regions of the input space with a high density of data points If data points are few and far between in the vicinity of the query point then LLR will not be very effective if the underlying function we are trying to model is truly non linear The winGamma User Guide BUILDING AND TESTING A MODEL Version 18 Jan 2002 It may seem odd that although winGamma is all about constructing smooth models the global function produced by patching together many LLR predictions in general is not even continuous However as the number of data points increases the global function produced by LLR will converge rapidly to the unknown
75. nt If the final target application is a real time system neural networks offer the advantage that they can be implemented in hardware How should I choose between local linear regression and dynamic local linear regression For a model to adapt it must be dynamic Every data row vector seen by a dynamic LLR model will be added to the model but of course eventually the model becomes memory hungry and starts to slow down At this point the model will have to be pruned If the phenomena that you are trying 74 The winGamma User Guide APPENDIX VI Frequently asked questions Version 18 January 2002 to model is likely to be fixed then a static model is best If the underlying dynamics themselves might be changing e g the stock market then a dynamic model is more sensible How should I choose between the Backpropagation Conjugate Gradient Descent and BFGS neural net algorithms Backpropagation is the original feedforward neural network training algorithm It is reasonably effective on simple problems but only makes use of the first differentials of the error surface in weight space Therefore backpropagation can take longer to train than other more sophisticated neural training algorithms and may fail to converge to the target MSError derived by the Gamma test at all But compared to more recent algorithms backpropagation is inexpensive on memory CGD offer some improvements over BP at the cost of extra memory BFGS uses the second differe
76. o change Whether normalization is a good or bad idea depends largely on the circumstances If input variables are incompatible then it is probably a good idea to normalize Normalization of just the input values will not change the asymptotic Gamma statistic or Vratio provided we imagine that as the number of data points becomes large we also increase pmax by a suitable constant factor but a good scaling will cause the M test to converge more rapidly to the asymptote so improving the accuracy of the noise estimate for a given amount of data A good scaling can also improve the accuracy of a model constructed using a fixed amount of data The effect of masking is all or none and it may be better to apply a suitable weight to each input variable For example it is a general observation regarding near neighbour classifiers that they perform well given the right weighting of inputs but that at present there are no general techniques for finding such weightings However if weights are applied then of course the data must NOT then be renormalised 2 16 Projects A Project is the collection of all Experiments performed on a given data set A given Project is determined initially once the data set for analysis is defined At Project creation time the number of inputs and outputs to a time series have to be set and options to normalise or scale the data There is also an option to generate a parallel moving average and or difference series along
77. of Figure 2 25 is intended to show the relative stability of the Gamma test result In this case it really fails to do so because the order of the data should really be randomised since it is very time periodic Even so when we examine the vertical scale of Figure 2 25 we see that the relative variation is not very large We see shall later in Chapter III how to take the results of this analysis and build and test models using the solar csv file 2 13 Analysing Time Series data 2 13 1 The DH 34 5000 asc data Delayed Henon Map ji j Figure 2 26 The first 100 points of the Figure 2 27 The return map x delayed Henon map time series delayed Henon map xp for the n 1 This Time Series data was generated by a process very similar to the Henon map except that where the current value of the Henon map time series depends the last two values of the series for the Delayed Henon map the current value is determined by the values three and four steps in the past This changes things in a number of respects The plot of the time series is given in Figure 2 26 and Figure 2 27 shows the return map for x Xp which is analogous to Figure 1 12 We observe that this distribution looks quite different 42 The winGamma User Guide PERFORMING AN ANALYSIS Version 18 Jan 2002 We now proceed through the steps outlined in section 2 1 1 for Time Series 1 Load the data and do not initially normalise it is a single time series
78. oints in steps of 10 from Sin500 asc 38 The winGamma User Guide PERFORMING AN ANALYSIS Version 18 Jan 2002 2 12 3 The solar csv data The data considered in this 2 input 1 output file relates to the generation of electrical power by an array of solar cells The inputs are a measure of light intensity South Plane Irradiance to a precision of 0 01 kW m and the current temperature in degrees C to a precision of 0 5 C Figure 2 19 Plot against time of the irradiance and temperature from the training data file solar csv The output is the voltage inverter AC power output measured to a precision of 30 01 kW The file consists of these values sampled every minute Figure 2 19 illustrates the graphs of the two inputs and the output against time position in the file We note that at low Irradiance the recorded power output values are irregular and sometimes negative This is a result of the fact that intelligent circuits are attempting to determine whether or not to initialise the system as the sun rises or sets The effect is to produce noise on the output power at low Irradiance levels We are just using the data as an example but if one wanted to use the data to build a really accurate model obviously one should filter out the data having low or zero Irradiance 7 Data for this example were provided by Newcastle Photovoltaics Applications Centre at the University of Northumbria at Newcastle UK These data were collected as p
79. om inputs to outputs using only the data provided Both the inputs and outputs should be continuous real variables from some bounded range The software will be much less effective if some of the input or output variables take only categorical values e g 0 or 1 The underlying function is presumed smooth and this means bounded first and second derivatives If the unknown function has regions of very high curvature it will be much harder to produce an accurate predictive model Itis also assumed that the noise variance on each of the outputs is bounded and independent of the input values If the independence condition is false this is not necessarily fatal the Gamma test will return an average noise variance over the whole input space Subject to these conditions winGamma can be applied to a wide variety of non linear modelling problems It is particularly useful in the research and design of non linear control systems 1 2 Loading data files winGamma can analyse two basic types of numeric data files Input Output data where each column corresponds to an input or an output and Time Series data where each column corresponds to a particular time series and successive rows represents successive values in time for each series Note all data files must contain only numerical data arranged in one of the allowed formats For more details of data file formats see Appendix II To load a data file launch the application from the Start menu Click on
80. ore but which are essential if the product is ever to enjoy widespread use Dr Nicola M Pearsall who kindly gave permission for us to use the example data in solar csv Data for this example were provided by Newcastle Photovoltaics Applications Centre at the University of Northumbria at Newcastle UK These data were collected as part of a project with funding from the European Commission THERMIE Programme and the UK Department of Trade and Industry To all of you we express our thanks and hopes that winGamma will make a contribution worthy of your efforts The winGamma User Guide CONTENTS CHAPTER I Getting Started 2iih ob RR DEG R RE Pxe AREE XAR RPIX EXAY 9 hol IntrGdUCloH cook Ee Eius bee tie ow SEXES EI e eX 9 1 1 1 The Purpose of the Software 24 4 6 lu eL Vp RE Ex 9 1 1 2 The range of applicability sre IR ReRE PREIS 11 1 2 1929 data files Ao eder aa Ate he ke Mi tue d oM hong Cual n 11 1 2 1 Comma separated variable csv files from spreadsheets 11 1 2 2 Input Output data in standard format asc files 12 1 2 3 Time Series data in standard format asc files llis 13 1 2 4 Partitioning the data xus peu ooo ER S expeh E SEXES ee koe 13 1 3 A first experiment 22 63 pl Xp xe RpEREPST p Een d prie 13 1 3 1 Interpreting tlie results 5 2 ee eSI ELESEV e Bees 14 1 4 The basic controls of winGamma 23 acea uu EE EX XS 17 gt Two simiple examples 22e ukek oti ex aed sete RSRERILEMESZem
81. ot of the FTSE close data 44 The winGamma User Guide PERFORMING AN ANALYSIS Version 18 Jan 2002 If we perform a full embedding on the last 20 weeks on a convenient Unix station using pmax 10 we obtain the histogram in Figure 2 33 This has the characteristic Gaussian shape of mainly statistically determined data From the table of the best embeddings we select the embedding which gives the smallest positive gamma This is 11011100110000111101 Handy Tip Typing in long masks can be error prone and tedious There is no need to do this A mask can be copied to the clipboard and pasted in whenever required using a right click on the mouse Results Visualiser Select output Output 1 v Custom Chatt X Series Unique Data Pc Y Unique Data Points v Gamm 1 600 4 Frequency 140000 lzo000 looo000 0000 50000 40000 0000 100 120140 160 180 200 220 240 260 260 300 320 340 360 380 400 420 440 460 480 Unique Data Points 000 4000 6000 000 10000 Figure 2 32 The M test graph for the FTSE Figure 2 33 The frequency histogram for data using the best embedding of length 20 embeddings of length 20 using the FTSE data This choice of embedding should be treated with some caution The M test graph of Figure 2 32 pmax 10 Randomised 3 repetitions shows that the estimated gamma values have not yet stabilised M is not sufficiently large so the error in estimating the Gamma statistic for any particul
82. perform the analysis on a subset of the data If the problems continue you need more RAM In most cases 64 Mgbytes is sufficient for any reasonable data set At least 50 Mbytes remaining hard disk space Operating system Windows 95 or 98 or Windows NT4 0 were the original development targets but we have so far observed no problems with later versions of Windows operating systems Licenses for a script file driven UNIX version of the Gamma Test software may be available by special arrangement Installation The winGamma User Guide APPENDIX I General Information Version 18 January 2002 Beta release At present simply copy all files in the winGamma directory into a convenient directory on your hard disk If you experience problems getting the help system to work you may have an older version of Explorer To update run the file hhupd exe V I release Place CD in drive Follow install instructions from screen List of files and directory structure after installation DIR DIR Program and associated files BORLNDMM DLL CP3240MT DLL hhupd exe Run to update HTML if problems with help occur Tee4C bpl vcl35 bpl winGamma chm winGamma exe winGammaBaseComponents bpl winGammaComponents bpl Directory of C WinGamma 11 20 98 12 28p lt DIR gt Data 10 30 98 02 37p lt DIR gt TestFiles 02 09 98 02 00a 29 952 02 09 98 02 00a 996 872 11 05 98 06 28p 471 840 10 24 98 04 01a 420 864 02 09 98 02 00a 1 455 736 02 22
83. port Choose Save as type 62 The winGamma User Guide BUILDING AND TESTING A MODEL Version 18 Jan 2002 Step 3 Set to Excel Macro mac Enter directory and file name Export model Setting up the data and model in Excel Start up Excel Load data from test csv Save the file as an Excel Workbook Right click on worksheet tab test at the bottom left corner Select Insert Select MS Excel 4 0 Macro Hit OK This now opens a macro sheet Load test mac into Notepad select all text and copy Paste text into macro sheet in Excel in cell A1 Highlight column A Do InsertiName Define In the macro box set to Function Set name to model Hit OK Now when in the macro sheet with column A highlighted you should see model in the top left name box Switch back using the tabs at the bottom to the test worksheet Enter heading model in cell F1 In cell F2 type model A2 C2 no quotes and press Return You should now see the model output value in cell F2 as compared to the actual output in cell D2 Select cell F2 and copy Highlight the range of cells from F3 to F278 and paste You should now have all the model predicted values for each row 63 APPENDIX I General Information Shipping list Compact disc 2 This manual 3 The gamma test and how to use it a practitioners guide Hardware requirements This software is PC based and normal minimum requirem
84. pulation to do from scratch in Excel Notice that although we set out with the intention of trying to model the time series Target we have created a file in which every time series has an out put that we can model You can delete these extra outputs in Step 6 if you wish Youcan begin immediately with experiments on this file but since not all these inputs may be needed for the model you can also proceed as follows Step 5 Use some other software such as Mathematica to perform data analysis on the last file using tools not yet provided by winGamma For example one useful analysis tool is to take the average lagged correlations of successive differences of the target time series with the successive differences of all the time series e g for lags from 1 to 8 in the above example This tool is available as DeltaCorrelation in the Mathematica suite provided with winGamma We may choose to take only those lags which have the largest absolute lagged delta correlation This may suggest that some columns could be deleted from the csv file we have produced Step 6 Load the csv file into Excel and delete the columns which have been selected as unlikely to be useful Re save the result as a csv file and proceed as if from the end of Step 4 with winGamma analysis on the resulting file Further inputs may be set to zero in the mask as a result of winGamma analysis Notice that when you reload this file into winGamma it will be treated as an
85. r pmax 10 for Ran500 asc Ran500 asc Analysis Manager B Results Visualiser Select output Output 1 X Select output Output 1 zl Custom Chart i X Series Angle Histogram Position in List zl d Position in List v Gamma i tae tt i tt i 1 Primary Y Series 1a dg Gamma z a x 8 E 5 5 8 5 2 3 4 5 8 7 8 9 10 11 12 13 14 15 16 17 18 19 20 Position in List 80 80 70 60 50 40 30 20 10 5 0 5 101520253035 40 4550 556065 70758085 Figure 2 11 Angle histogram for Ran500 asc Figure 2 12 Moving window Gamma test pmax 10 pmax 10 on 300 points in steps of 10 from Ran500 asc 36 The winGamma User Guide PERFORMING AN ANALYSIS Version 18 Jan 2002 2 12 2 The Sin500 asc data If we run a simple Gamma test Table 2 2 The Gamma test result pmax 10 for unscaled and scaled data with pmax 10 near neighbours the file Sin500 asc we obtain the results in Table 2 2 0 07335 0 03190 0 71122 4 0386 The estimated Gamma statistic 0 07335 indicates a moderate noise level as does Vratio 0 12762 The regression line on the scaled data with slope A 4 0386 indicates definitely non linear data However if we pull up the Gradient plot we see that it is still Lower 95 Confidence l highly variable so one should Mask remaining entries not have too much confidence in this observation Taken together these results indicate noisy but manageable non lin
86. rner of many of the graphics windows 1 8 Customising the file and project directories To customise the locations of data files and project files discussed in Chapter IT pull down the Options menu and click on Customize You can modify the number of data files and project files kept in the history in is usually best to set these to their maximum of 9 Now under data files click on Modify and select the directory that should first appear when the process of loading a data file is initiated When the desired location has been selected click on OK Go though a similar procedure to locate the project directory If you wish the windows settings to be saved each time winGamma is closed down then check the appropriate box Finally click on Apply and exit the program 23 The winGamma User Guide GETTING STARTED Version 18 Jan 2002 24 CHAPTER II Performing an analysis 2 1 Introduction An Experiment is a particular type of calculation performed on the analysis data A new experiment is started by highlighting the type of experiment required and then selecting New in the Analysis Manager window If we want to perform the same calculation but with different parameters e g the number of nearest neighbours or a new method e g M test then a new experiment is started In this chapter we discuss each type of Experiment how to set the parameters and how to interpret the results Each Experiment is discussed using an example and illustra
87. rom a continuous range If many inputs are categorical it is also possible to get a negative Gamma statistic How should I choose the right number of inputs for a Time Series Initially set the number of inputs large but reasonable in the context of the data Then do an Increasing embedding This will compute successive Gamma statistics based on one input the historically most recent sample of the time series rightmost on the mask then on two inputs the two most recent samples and so on up to the maximum number of inputs you have selected The minimum Gamma statistic obtained will determine an upper bound for the maximum number of inputs it is useful to consider An optimum for the number of near neighbours used in the Gamma test should now be obtained Then the maximum number of inputs can be checked again using that number of near neighbours in the Gamma test If the maximum number of inputs changes then the optimum number of near neighbours should be checked again Finally using the best maximal number of inputs a check for the best embedding can be run this may cause some inputs to be discarded How should I choose a method for establishing an optimal embedding mask The best method for choosing a mask on the inputs is Full embedding The problems come with this method when the run times required become too long Runtime is a function of the input dimensionality the number of inputs m the number of nearest neighbours pm
88. s 2 sigmoidal act scaleFactor 1 l e act Temperature where act is the activation weighted sum of inputs scaleFactor 1 5 and Temperature 0 8333 To speed up neural computations this function is implemented in winGamma as a fine grained look up table whereas for feedforward computations when the weights are loaded into other software it can be implemented directly as a function This may cause very small differences in neural output calculations using the same weights outside winGamma 3 9 2 NetReader NetReader nb is a Mathematica program supplied with winGamma which can read the neural network weights saved from winGamma and implement the neural network for feedforward testing Which type of network training was used in the creation of the weights is automatically identified from the weights file 3 9 3 Exporting and using Neural network models in Excel After winGamma has built a neural network model it may be exported as an Excel Macro and used directly in Excel This facility is not currently available for LLR models We illustrate this process using an example Step 1 Build a model In winGamma load data file Sun280 asc this is a single time series file Transform the data to 3 inputs 1 output Export transformed data as test csv Perform Gamma analysis Train neural network model on the transformed data Step 2 Export the model Right click on Model Select Ex
89. se tabs with those produced for the noisy Sin data in section 1 5 1 1 We shall examine these tools more fully in Chapter II 1 6 Linear models winGamma is a non linear modelling tool and makes very few assumptions about the nature of the model Because of this fact it generally needs far more data than parametric analysis where the model is presumed to have a particular form If it is safe to assume that the model is linear then a simple linear regression model should be built and tested using some other standard software e g Mathematica has very good linear regression facilities If you know nothing at all about the data being analysed it is always a good idea to check the linear regression model before using winGamma If the data is fundamentally linear then winGamma will perform quite well using local linear regression However winGamma will make less efficient use of the data available than global linear regression 1 7 Exporting results for use by other software Data produced by winGamma is either Graphics or data such as predictions Data Files can be exported in 1 Mathematica compatible format e g s are embedded to format lists and arrays 2 Excel and spreadsheet compatible comma separated variable csv format These Export functions are available as an option under Edit in the main winGamma parent window a right click on the mouse button in the appropriate context or by clicking the gt tab in the top left co
90. side a time series As successive Experiments are completed the parameter settings for each Experiment and the results are added to the Project which can be saved as a file and reloaded at a later date Thus there is no need to repeat the same experiment which may have taken a while to compute All distance functions are equivalent to within a constant But rescaling changes specific near neighbour relationships 50 The winGamma User Guide PERFORMING AN ANALYSIS Version 18 Jan 2002 Handy Tip The Project file gpr contains the path of the data file used in the project in fact the data file name and path are the first item in the Project file Project files are in DOS ASCII format and can be edited as long as the file is saved in the same format after editing This can be useful to know if the data has been moved or renamed or you have moved both the Project file and the data file to another system where the path is different One way to handle this is to just edit the path in the Project file However if winGamma cannot find the data file associated with a project it will ask you to Browse for the file and you can indicate the new location It is important to select the right file which must not have been altered 51 CHAPTER III Building and testing a model 3 1 Introduction We now assume that you have analysed the data and decided which inputs and how much data to use To actually build the model winGamma offers several tec
91. sion 3 Transform the data set to reset the maximum number of inputs to the largest number from the Increasing Embedding Experiment which still gives a comparatively small Gamma statistic 4 Run a M test to check the stability of the Gamma statistic If the M test produces a stable asymptote decide if the noise variance is likely to be 7ero arbitrarily good models possible with enough high precision data Or Non zero not much point is using more data than necessary to give a model which predicts at the Gamma statistic level On this basis decide how much data is likely to be needed to build a model 5 Can we get a better Gamma statistic by discarding some of the input To answer this question run a Full Embedding if the number of inputs is small enough to allow this say 10 15 Otherwise try the heuristic search techniques ending up with a long GA run 6 If a better embedding is found then repeat steps 4 5 and 6 to refine those conclusions 7 Refine the number of near neighbours for the final estimate of the Gamma statistic using an Increasing Near neighbours test 2 2 The Gamma test 26 The winGamma User Guide PERFORMING AN ANALYSIS Version 18 Jan 2002 This finds the Gamma statistic and other relevant measures These are principally the Gradient the Vratio and the Standard Error as described in Chapter I Once the inputs have been determined either with preliminary Gamma tests or because these are set by t
92. suffix asc Data files may be created using Excel as csv files and imported into winGamma Data files for winGamma are in two basic formats e Times series data Example a single time Series 0 0262 0 0575 0 0837 0 1203 0 1883 0 3033 0 1517 etc Each number followed by a carriage return linefeed Example multiple time Series it is the responsibility of the user to prepare the data so that fields referring to the same time are on the correct line most recent data is last 0 0262 1000 26 0 0575 1031 78 0 0837 1037 86 0 1203 1038 567 0 1883 1040 810 0 3033 1100 721 0 1517 1027 851 Each number followed by one or more spaces The last number on a line followed by a carriage return linefeed There must be the same number of data fields on each row e Input Output data Example a 4 input l output file 0 36368593157164 0 3304959949667 0 21811098544356 0 2093396 1443087 0 0220710621963 0 00591 105325917 0 9085902611647 0 19548859472561 0 34015487882487 0 0064356217878 0 86221883819100 0 5929180658183 0 36843151702318 0 89277930056707 0 6617039028787 0 59877814813365 0 9562473549851 0 25582643936911 0 97996127233012 0 4810764303063 0 13712162278232 0 9035299186427 0 29916358157799 0 22014139763247 0 7734356912106 0 42696607632396 0 4827254329784 0 98919821679839 0 20449324659299 0 5789449769352 etc The winGamma User Guide APPENDIX II Data file formats Version 18 January 2002 Each number fol
93. t graph for Sin500 asc Note the relatively stable asymptote 31 Figure 2 6 The Ran500 asc output plotted against the position in the file 34 Figure 2 7 Increasing near neighbours 3 30 on Ran500 asc Gamma SE 36 Figure 2 8 M test pmax 10 on Ran500 asc 1 eee eee 36 Figure 2 9 Scatterplot and regression line pmax 10 for Ran500 asc 36 Figure 2 10 3D Histogram pmax 10 for Ran500 asc 6 eee 36 Figure 2 11 Angle histogram for Ran500 asc pmax 210 llle 36 Figure 2 12 Moving window Gamma test pmax 10 on 300 points in steps of 10 from RANDOO OSC i i p sve NAE SE XGA TES GA EGO RUE RS RR da es REA MARE 36 Figure 2 13 Increasing near neighbours 3 30 on Sin500 asc Gamma SE 38 Figure 2 14 M test pmax 17 on Sin500 asc llle 38 Figure 2 15 Scatterplot and regression line pmax 17 for Sin500 asc 38 Figure 2 16 3D Histogram pmax 17 for Sin500 asc 1 eee 38 Figure 2 17 Angle histogram for Sin500 asc pmax 17 1 1 00 eee eee 38 Figure 2 18 Moving window Gamma test pmax 17 on 300 points in steps of 10 from SIHOUO USE 2 areas iter eet de WUE ek et Ba Pa ee Rit he EE Mead use pus 38 Figure 2 19 Plot against time of the irradiance and temperature from the training data file ovis ACER 39 Figure 2 20 Increasing near neighbours 3 50 on solar csv Gamma SE 40 Figure 2 21 M test pmax 20 Randomised 2 repe
94. t might in model construction to choose a mask with a low Gradient which will correspond to a simpler model and the fitness due to the number of 1 s in the mask because shorter masks also mean simpler models The contribution of each of these terms is controlled by three weights Winrercepr W gradient Ad W engin according to the formula fitness mask W intercept W gradientFitness mask gradient Wren ulengthF itness mask AinterceptFitness mask The component fitness calculations are described below where Vratio mask and Gradient mask return the Vratio and Gradient as calculated by the Gamma test on the data set for mask outputrange is the range of the output and denotes absolute value interceptFitness mask 1 1 IO Vratio mask ifVratio mask lt 0 2 2 1 Vratio mask otherwise gradientFitness mask 1 1 gradient mask outputRange numofones mask lengthFitness mask length mask 712 APPENDIX VI Frequently asked questions Why is the Gamma statistic sometimes negative Sometimes the Standard Error the error obtained from the 0 y regression which is always stated when a Gamma result is obtained is large enough to account for a negative intercept by the regression line This is most likely to occur when the true asymptotic Gamma statistic is close to zero It can also happen when the data fails to fulfill the basic requirement that inputs and outputs are drawn f
95. t present these are the same for all series In Figure 1 2 we are choosing to use 5 previous values of every time series to predict the next 2 values for each of the time series Choosing more outputs will produce predictions further into the future The nature of things is such that the further we try to predict into the future the less accurate these predictions will be This is reflected in a higher Gamma statistic for more distant future predictions 1 2 2 Input Output data in standard format asc files Data Transformation Standard format for an Input Output file is DOS ASCII in the following form In each row the inputs are separated by spaces and the list of Data Settings inputs terminated by a comma The list of outputs then follows each separated by spaces The end of Data type Vector Function a row is signified by CR LF File data in standard Inputs Input Output form will be automatically Dutputs recognised as such At present the numbers in the Vectors file must be in simple decimal format The first decision to be made after specifying the file name is whether or not to Transform i e to scale or normalise the data To normalise check the appropriate box as indicated in Figure 1 3 Figure 1 3 The Normalise check box For a full discussion of the effects of scaling and whether or not to scale in any particular case see section 2 14 In an initial investigation it is usually a good idea to scale Input Ou
96. t value Mean squared error MSError If y i 1 lt i lt M is a set of values of an output and y 1 is a set of predictions for y i then the MSError of the predictions is given by M Y yx xf MSError i Mi 1 Standard Error SE This is the standard error about a regression line and is calculated as pmax SET Y qo TY n 2 i l where I 1 is the ith Gamma regression point value and T is their mean Over training describes the effect when we attempt to produce a model by exactly following the training data Consider the effect of trying to produce a model by drawing a line through every point in the noisy sine data in Figure 1 8 It would look nothing like a sine curve and if we asked this model to predict y for a particular value of x we should little faith in the prediction One of the main advantages of winGamma is that it gives us the necessary information to prevent over training before we begin to build a smooth model such as a neural network GA Fitness In order to better control the GA search it is useful to know how the GA fitness is calculated The overall fitness of a mask is composed of three parts corresponding to the fitness due to the intercept i e the actually Gamma statistic because mainly we want masks with small Gamma The winGamma User Guide APPENDIX V Definitions Version 18 January 2002 the fitness due to the Gradient because if we have enough data to estimate the Gradient accurately i
97. ted Gamma result Therefore it is important to highlight the required result in the Results window 2 3 The scatter plot and regression line The critical graph to look at first is the scatter plots and d p y p regression line see Figure 2 1 The scatter plot shows point pairs y where 6 is the squared distance of an input x from one of its near neighbours and y is one half of the squared distance between the two corresponding scalar output y values The points to which the regression line is fitted are calculated by finding the mean p of 6 and y p of y where p refers to the first nearest neighbour p 1 the second nearest neighbour p 22 and so on up to the maximum number of near neighbours pmax which has been set by the user A good regression line with points 6 p y p approaching 6 y 0 0 indicates that the scalar output values of input near neighbours are close If the regression line has a steep slope this indicates that the modelling function fthat we seek to approximate is liable to be quite difficult to construct and 27 The winGamma User Guide PERFORMING AN ANALYSIS Version 18 Jan 2002 a large number of data points M will be required If the line is almost horizontal the function is quite simple A particular feature to look for here is an empty wedge in the top left corner of the scatter plot If there are points in the top left corner it means that there are input points in the or
98. ted with screen shots 2 1 1 The user cycle The user cycle for a full analysis is not completely fixed and can be varied according to circumstances However the general approach can be summarised by the following steps Input Output files 1 Load the data and on the full data set if not exceedingly large do a simple Gamma test scaled and unscaled with the number of nearest neighbours set to the default of 10 If the data set is very large use a subset of the data for initial experiments 2 Run an ncreasing Near Neighbours test and use the minimum SE between say pmax 5 and pmax 50 to determine the most accurate Gamma statistic 3 Using the value for pmax determined in Step 2 run an M test to determine how stable the Gamma statistic is with increasing data set size 4 If the M test produces a stable asymptote decide if the noise variance 1s likely to be 7ero arbitrarily good models possible with enough high precision data Or Non zero not much point is using more data than necessary to give a model which predicts at the Gamma statistic level On this basis decide how much data is likely to be needed to build a model 5 Can we get a better Gamma statistic by discarding some of the input To answer this question run a Full Embedding if the number of inputs is small enough to allow this say 10 15 Otherwise try the heuristic search techniques such as Hill climbing or Sequential embedding see 2 7 2 2 7 4
99. th the embedding ee i X Series search In this case we see a somewhat 53 Primary Y Series Gamma y Overlay Y Series forem z 5 E Near Neighbours v Gamma and Standard Error irregular multimodal histogram there are only 15 possible embeddings 6 If we now fix on the embedding 1101 having a Gamma statistic of around 2 3228E 6 then we might next do an Increasing Near Neighbours Experiment to optimise the choice of near neighbours in estimating the Gamma statistic 34567 8 9 1011121314151617 1819 20 21 2223 24252627 28 2930 Near Neighbours Gamma Standard Error Results The result is shown in Figure 2 30 Figure 2 30 An Increasing Near Neighbours The minimum SE is obtained using pmax est on the embedding 1101 for the delayed 7 nearest neighbours and corresponds toa Henon map final Gamma statistic of 2 3228E 6 so that optimising the number of near neighbours hardly changed the Gamma statistic at all in this case Thus the final analysis of DH 34 5000 asc is that it is a noise free time series Using a few thousand points we should be able to construct a model capable of one step prediction with an estimated MSError of around 2 23E 6 2 13 2 The FTSE weekly closing price data The file FTSEcls asc contains the FTSE weekly closing price from 9 May 1988 26 January 1998 which gives 508 samples Figure 2 31 shows the time series over the full run of the data Time weeks Figure 2 31 A pl
100. the surface constructed from the data actually looks like Figure 3 9 This is a topographic plot of the surface in which lower power outputs are blue and higher power outputs are red We could regard our Whatlf graphs as cross sections using the 10 Temperature Irradiance Figure 3 9 A topographic plot of the solar csv data model of a surface which is very similar to this plot 59 The winGamma User Guide BUILDING AND TESTING A MODEL Version 18 Jan 2002 3 7 Example model construction and testing for DH 34 5000 asc In Chapter II we analysed this data and concluded that it represented a low or zero noise time series for which the current sample could be predicted accurately on the basis of between 4 and 7 previous samples The mask 1101 was identified as a good mask 1 Load the file DH 34 5000 asc 2 In the Time Series options specify 4 inputs and 50 outputs This gives 4952 samples for analysis We are going to build a local linear regression model and this needs a kd tree so it is necessary first run a Gamma test 3 In the Experiments tab highlight LIzzArz Gamma test Leave the number of oem ETE sl near neighbours set at the default of amp pss 10 In the Mask tab enter the maskas Model Tester 1101 Now click Execute KAN i Result We specified 50 outputs and so are trying to predict a maximum of 50 steps into the future If we just look at the one step prediction then the res
101. this is the most recent and obtains a Gamma statistic for this mask It progressively increases the number of bits set in the mask working from right to left performing a Gamma test for each new mask It runs to the maximum number of inputs and stops We can then examine the Gamma statistic for each mask The best embedding found will be the one whose Gamma statistic is closest to zero This is useful in a time series to discover the underlying embedding dimension as we saw in section 1 5 2 In the next sections we shall give example analyses using these various options 2 12 Analysing Input Output data 2 12 1 The Ran500 asc data We begin with a data set which is a type of worst output case in the sense that there is no smooth data model for this example The file Ran500 asc is is a 4 Input 1 Output file containing 500 I O pairs of completely random data generated using a uniform distribution in 1 1 via the Mathematica test file DataGen nb The output is actually pure noise having a true variance of 0 333333 A point plot of the output is given in Figure 2 6 i If we run a simple Gamma test with pmax 10 near neighbours we obtain the results in Table 2 1 The estimated Gamma statistic 0 31793 indicates Figure 2 6 The Ran500 asc output plotted a high noise level as does Vratio 0 97821 which against the position in the file is very close to one The regression line with slope A 0 0575 on scaled data is close to hor
102. titions on solar csv luus 40 Figure 2 22 Scatterplot and regression line zoomed pmax 20 for solar csv 40 Figure 2 23 3D Histogram pmax 20 for solar csv lees 40 Figure 2 24 Angle histogram for solar csv pmax 220 0 0 eee ce eee 40 Figure 2 25 Moving window Gamma test pmax 20 on 8400 points in steps of 100 from NOVA gro y MN CETTE 40 Figure 2 26 The first 100 points of the delayed Henon map time series 42 Figure 2 27 The return map x Xp for the delayed Henon map s sss 42 Figure 2 28 The result of an Increasing Embedding for the delayed Henon map 43 Figure 2 29 The M test graph pmax 210 number of inputs 8 for the delayed Henon map EUR I TD M KE EE 43 Figure 2 30 An Increasing Near Neighbours test on the embedding 1101 for the delayed Henon TUBOS cec riesce pete qu ud sposi ameet bars ros eta A ed actes acra ee 44 Figure 2 31 A plot of the FTSE close data 0 eee eee eee eee 44 Figure 2 32 The M test graph for the FTSE data using the best embedding of length 20 45 Figure 2 33 The frequency histogram for embeddings of length 20 using the FTSE data 45 Figure 2 34 Plot of the sunspots data file Sun280 asc llle 46 Figure 2 35 The M test graph for the sunspot data data using the best embedding of length 15 prede AE EN Pan dae du AER SE ARCS eae Re Pda O dad s 47 Figure 2 36 The frequency histogram of all embeddings of length 15 using the
103. tput or multiple Time Series data 12 The winGamma User Guide GETTING STARTED Version 18 Jan 2002 1 2 3 Time Series data in standard format asc files Standard format for a Time Series file is DOS ASCII in the following form Each column represents an individual time series The rows represent values for each of the time series successive rows being successive values in time Within a row each numeric value is separated by spaces The end of a row is signified by CR LF 1 2 4 Partitioning the data Select proportion of data set for analysis Start 1 10578 End 1 Cancel Figure 1 4 Selecting a proportion of the data for initial analysis It is sometimes convenient to perform the initial analysis on a subset of the whole data file This could happen for example where the data set was very large Therefore winGamma will next ask the user to select the proportion of the data which should currently be used for analysis see Figure 1 4 We can later separate training and test data 1 3 A first experiment Load the 2 input l output data file solar csv and select column 3 as output Initially do not normalise Select all the data for analysis there are 10578 data points in the file After the data has been successfully loaded winGamma displays the main screens as in Figure 1 5 The Experiments window in the Analysis Manager shows the different kinds of data analysis that can be performed We shall discuss the meaning of th
104. ult for the first output is a Gamma statistic of 6 089E 5 with a SE of 4 3 1 1 8E 5 3 000 3 020 3 040 3 060 3 080 3 100 3 120 3 140 3 160 3 180 3 200 Actual Predicted Error j 4 For output 1 select Model Our Figure 3 10 A test of the LLR model on the data previous experiments suggest that lt et DH 34 5000 asc blue predicted green actual about 3000 data points are need to red error obtain a good model a fact confirmed by a M test for this embedding which curiously gives small negative results increasing towards zero Select a local linear regression model with 10 nearest neighbours and set the Mask to 1101 Results The MSError over the test set 3000 The MSError of the test set 3000 3200 is 8 2303E 6 The graph of predictions actual values and errors is shown in Figure 3 10 3 7 1 How the prediction quality degrades into the future 60 The winGamma User Guide BUILDING AND TESTING A MODEL Version 18 Jan 2002 Results From the result of Step 3 in last experiment we actually got 50 output Gamma statistic results The graph of Gamma against the number of steps ahead is shown in Figure 3 11 Here we see an exponential rise in the error of prediction which is typical of a chaotic process We conclude that the Time Series data is of a low zero noise smooth process which 1s chaotic
105. um amount of time to spend on trying to attain the MSError goal This is shown in Figure 3 2 55 The winGamma User Guide BUILDING AND TESTING A MODEL Version 18 Jan 2002 Handy Tip Backpropagation along with most other processes in winGamma can be paused resumed or terminated using the buttons on the top level menu Terminating an operation does not necessarily loose everything Any results already calculated will be displayed and in the case of neural net training the model created so far will be retained Figure 3 3 shows the Analysis Manager during backpropagation training Note that the graphical window can be zoomed and moved using the left and right mouse buttons Because the number of layers the number of hidden units and the slope of the sigmoidal are fixed limiting the size of the weights also limits the magnitudes of the partial derivatives of the neural network as a function of its inputs Thus if the unknown function to be approximated has regions of high curvature the training algorithm with regularisation may find it difficult to obtain the desired approximation We can get some idea if this is likely to be the case by examining if the Gradient returned by the Gamma test is unusually large These parameters may be left at default until fine tuning of the model is required 3 4 Conjugate gradient descent This a variation and improvement on two layer vanilla backpropagation itis generally more effective but requires more memor
106. winGamma to scientific data However since the data in this file only runs over about one week we do not consider this extra complication here If we run a quick Gamma test on the full data set with pmax 10 near neighbours we get the results of Table 1 1 in Chapter I The unscaled Gamma statistic of 0 020761 seems high but in view of the output range approx 0 30 is actually quite good A better measure is the V 0 000760 defined as the ratio Gamma Var output which is low and shows that the output is highly predictable from the inputs Because the data clearly falls into two distinct classes day and night we should be aware that representative training and test data should include both types The point to grasp here is that although the time series data varies from moment to moment as clouds obscure the sun the relationship between sunlight input at a given temperature and power output is a smooth almost linear model The next step in a more careful Table 2 3 The results of the Gamma test pmax 20 for unscaled and analysis is to run an Increasing scaled data from solar csv near Neighbours test This will 77 Unscaed Sealed give us some idea of the best 0 020328 0 000221 pmax to choose to give the most 0 250261 0 230184 ies e e b Standard Error 0 002051 3 095267E 5 Lea e dd or 0 000744 0 000884 the increasing near neighbours Near Neighbours test run for pmax 3 to 50 We Stat oo t 1 0
107. wing sections 2 4 Increasing near neighbours This experiment shows how the Gamma statistic and the other results returned by the Gamma test varies with the number of near neighbours used to compute it It is used to get some idea of how accurate the Gamma statistic is liable to be If we perform this experiment and use the graphing facility to plot the Gamma statistic and the SE against the number of near neighbours by examining the graphs together we can usually see which choice for the number of near neighbours is likely to produce the most accurate estimate For example in Figure 2 4 produced from Sin500 asc we see that the SE first increases and then for a while plateaus before eventually beginning to steadily increase The range of the plateau is roughly between 7 27 near neighbours and it minimises at around pmax 17 with a Gamma statistic slightly larger than 0 074 which we know from the way the data was constructed is close to the correct value 1 Results Visualiser Select output Output 1 Custom Chart X Series Near Neighbour Primary Y Series Gamma Overlay Y Series Standard Error 12131415 1617 18 19202 Near Neighbours Gamma Standard Error Figure 2 4 The variation of Gamma and SE as the number of near neighbours increases 30 The winGamma User Guide PERFORMING AN ANALYSIS Version 18 Jan 2002 We also note that the Gamma statistic is reasonably stable in the same range It is
108. y The procedures for set up are very similar 3 5 BFGS neural network Probably the fastest and most efficient neural network training algorithm offered by winGamma is a modified version of the Broyden Fletcher Goldfarb Shanno learning algorithm This algorithm uses second differences and is sometimes degraded by very noisy data but generally it is probably best to use this option first when trying to produce a neural model 3 6 Example model construction and testing for solar asc We return to the example solar panel data we analysed in 2 12 3 Using the first 8400 data and scaling we initially build a local linear regression model with k 20 We then test this model on the remaining points in the data file 3 6 1 Building and testing a LLR model We initially build a LLR model using the first 8400 points the results of our earlier analysis suggest that slightly more points are required for a really good model 1 Load solar csv Do not normalise and use all the data for analysis Execute a simple Gamma test these steps are described in 1 3 and 2 12 3 Handy Tip winGamma requires that at least a simple Gamma test Experiment be conducted before any attempt to build a LLR model a kd tree is required 56 The winGamma User Guide BUILDING AND TESTING A MODEL Version 18 Jan 2002 2 After the Gamma test results appear click on Model in the Analysis Manager 3 Select the training set as 1 8400 4 In the Modelling Editor leave th

Download Pdf Manuals

image

Related Search

Related Contents

Westinghouse 42-Inch Instruction Manual  VENTS TT PRO 150 Use and Care Manual  Lexibook DJ028SP compact camera  Bushnell 78-9500 User's Manual  WARNING: - Umarex USA  u550 manual del usuario  American Standard Twin Ell 8888.120 User's Manual  Les annonces Mode d`emploi Réservées aux membres SCBA à jour  カタログ(PDF) - 株式会社エフ・エム・アイ  5DII-P manual  

Copyright © All rights reserved.
Failed to retrieve file