Computentp, Neural Nets and MCLIMITS
This page has been substantially rewritten (and remains a work in progress) to focus just on the information required for a successful run of the Computentp and Neural Net package, to deliver exclusions. For information on results obtained using inputs created in v12 of athena, please refer to the archive.
Project Aims
This project aims to document the use of an Artificial Neural Network (ANN) system and fitting software for the analysis of data from inclusive Higgs searches at ATLAS involving a lepton trigger and Higgs decay to b+bbar. This will use input from the Computentp software, designed to automate the weighting of the input files as required.
Tools
ANN :- This is a kind of algorithm with a structure consisting of "neurons" organised in a sequence of layers. The most common type, which is used here, is the Multi-Layer Perceptron (MLP), which comprises three kinds of layer. The input neurons are activated by a set trigger, and once activated they pass data on to a further set of "hidden" neurons (which can in principle be organised into any number of layers, but most frequently one or two - in our case one), and finally the processed data is forwarded to the output neurons.
The key feature of a neural network is its ability to be "trained" to recognise patterns in data, allowing high efficiency algorithms to be developed with relative ease. This training is typically done with sample data which has been generated artificially, resulting in an algorithm that is very effective at recognising certain patterns in data sets. The only shortcoming is the danger of "over-training" an ANN, meaning that it becomes overly discriminating and searches across a narrower range of patterns than is desired (one countermeasure is to add extra noise to training data).
Computentp :- Simply running the code as above will result in less than optimal Neural Net training. The training procedure requires equal numbers of events from signal and from background (in this case it results in half of the signal events being used in training, half for testing). However, the above code will take events from the background signal samples in proportion to the file sizes - these result in proportions not quite in accordance with physical ratios. As the Neural Net weights results according to information about the cross-section of the process and so on stored in the tree, the final result is that while the outputs are weighted in a physical fashion, the Net is not trained to the same ratios, and so is not optimally trained. To solve this problem, Computentp is used to mix together all background and signal samples., and assign
TrainWeights to them, so that the events are weighted correctly for the Net's training.
Preparing samples for the Neural Net
Samples are produced for the Neural Net from AODs - results have previously been obtained for MC samples derived from v12 and v15 of athena. Current efforts are directed toward debugging the v15 results, and then upgrading to v16 input. The inputs are created from AODs using the
TtHHbbDPDBasedAnalysis package (currently 00-04-18 and its branches are for v15, 00-04-19 is for v16).
Current samples in use
Input data and cross-sections
These cross-sections are for the overall process, at √s = 7
TeV.
The ttH sample cross-sections are provided for the overall process - the MC is divided into two samples with W+ and W- independent of one another. These two samples are merged before being put through the ANN.
The tt samples were initially generated to produce the equivalent of 75fb-1 of data, based on the LO cross-sections. Taking into account the k-factor of 1.84, this means that now all samples simulate 40.8fb-1 of data. These samples have also had a generator-level filter applied - most events (especially for tt+0j) are of no interest to us, so we don't want to fill up disk-space with them, so we apply filters based on the numbers of jets etc. The Filter Efficiency is the fraction of events that pass from the general sample into the final simulated sample. To clarify how all the numbers hang together, consider the case of tt+0j. We have simulated 66,911 events - as said above, this corresponds to 40.8fb-1 of data. We have a Filter Efficiency of 0.06774, so the full number of events that a complete semi-leptonic event would be comes to 987,762 events in 40fb-1. Divide this by 40 to get the number of events in 1fb-1 (i.e. the cross-section), and you get 24,694 events per fb-1. Our starting point for our cross-section is 13.18, with a k-factor of 1.84, which gives a cross-section of 24.25 - so all the numbers compare with each other pretty favourably. This of course makes getting from the number of sensible state events to the number expected per fb-1 rather easy - simply divide by 40.8.... You'll notice that the cross-section includes all the branching ratios already, so we don't need to worry about that.
**IMPORTANT** The Filter Efficiency for these samples was calculated based on a no-pileup sample. The filter is generator level, and one of the things it will cut an event for is not enough jets. However, pileup adds jets, but these are added well after the filter. The net result is that a number of events that failed the filter would have passed, had the pileup been added earlier in the process. This means the filter efficiency (and thus the cross-sections) are incorrect, by a yet to determined amount....
For the other samples, however, we do need to worry about branching ratios - the quoted initial cross-section includes all final states, so we need to apply branching ratios to the cross-section to reduce it down, so that it reflects the sample we've generated. We then subsequently need to reduce the cross-section further so that it reflects the number of sensible states.
These cross-sections and branching ratios are correct as of 8 Feb 2011. qq→ttbb (EWK) is currently not being used, thanks to a bug in the production of the MC
Number of events surviving preselection, weights and TrainWeights
(See later in the TWiki for an explanation of weights and
TrainWeights.) This table will be completed with all the relevant weights and
TrainWeights at a later date - these values are to be compared to the output from Computentp to ensure everything is working as intended, and are calculated for the sensible cross-sections/events. (A quick check of the
TrainWeight is to multiply the number so events of each background by their
TrainWeight and sum them - by design, this should equal the number of entries in the ttH sample.)
Issues still to be resolved
1. In share/TtHHbbSetups.py:
include( "AtlasGeoModel/SetGeometryVersion.py" )
include( "AtlasGeoModel/GeoModelInit.py" )
Athena warns that both of these files are obsolete - this does not lead to an ERROR or WARNING, but nonetheless needs to be looked at.
2. In src/PreselectLeptons.cxx:
const EMShower* shower;
if ((*elecItr)->author() != egammaParameters::AuthorUnknown){
trackIsolationEnergy20 = m_trackIsolationTool->trackIsolationEnergy((*elecItr)->trackParticle(),0.2);
trackIsolationEnergy30 = m_trackIsolationTool->trackIsolationEnergy((*elecItr)->trackParticle(),0.3);
trackIsolationEnergy40 = m_trackIsolationTool->trackIsolationEnergy((*elecItr)->trackParticle(),0.4);
shower = (*elecItr)->detail<EMShower>(m_showerContainerName);
//CCT: shower - seems to have stopped working in 15.6.0.2 for data made with r838 (v15.3.1.6)
//was giving a seg fault as it tried to get etcone20 even if "shower" had not been successfully obtained!
if (!shower) {
mLog << MSG::WARNING << "Invalid EMShower pointer!" << endreq;
}else{
etcone20 = shower->etcone20()/pT;
mLog << MSG::INFO << "shower->etcone20() = " << shower->etcone20() << ", pT = " << pT << endreq;
}
}
The warning message appears an awful lot - also, don't think this influences the Net inputs, but still should be looked at - do we need to get anything from the shower?
3. In the jobOptions file we currently have:
PreselectLeptons.McEventInfoName = "MyEvent"
However, in the athena output we still have:
StoreGateSvc WARNING retrieve(const): No valid proxy for object McEventInfo of type EventInfo(CLID 2101)
Need to work out why this jobOption does not over-ride the default. It might itself be overridden by python/ttH_defaults.py - if this is the case, then a number of other settings are also over-ridden.
4. From the ye olde code, I noticed a 'placeholder' warning to check that W and top masses used in the sensible states are the same as in the generator (this hasn't been done yet). Also, perhaps we can tighten the sensible cut on the W mass? Seems rather wide at the mo (25GeV)....
Running the Neural Net
Overview of the process
-
Computentp takes in the original root files from ${ntuple_area}, and is steered based on steerComputentp.txt (created by genemflat_batch) . Computentp calculates the TrainWeight etc based on nGenForType etc, stored within the root files - no external hard-coding (I think).
-
There are two main outputs from Computentp:
-
computentp_output/ contains a file for each root file, containing all the variables for the tree - including weight and trainweight.
-
trees/ contains all the same information, in one large file.
-
The training runs on the Computentp output (the single large file), and produces the weights file (weights/) . It uses TrainWeight (branch inside Computentp output).
-
The templating is what actually produces the NN score plots. It uses one file per signal/background - uses the original root files, and the ANN weight files as produced by the training. Will calculate the scale-factor to apply to each event (also often referred to as the weight) based upon NGenForType etc, which are within the root files.
-
Using the MCLIMIT program, these ANN outputs are used to generate 1,000 pseudoexperiments (this number is set in the code, and can be adjusted if desired – the variable NPE in genemflat_batch).
-
For each pseudoexperiment we simulate a Background-only sample of ANN output, which is subjected to both systematic and statistical (Poisson) variations. The Poisson variations are applied twice – once to each individual background, and then again to the sum of the backgrounds. (A possible bug?)
-
For comparison to these null hypothesis datasets, for both signal and background an array of 10,000 templates is created, each of these templates being subjected to systematic and Poisson variation.
-
The pseudodata and the array are passed to the function cslimit, which uses them to calculate an exclusion for each pseudoexperimen.
-
Based on these 1,000 separate exclusions, we produce a final exclusion.
Setting up
Initial Setup
In the file genemflat_batch_Complete2.sh, you should correct the line:
#PBS -j oe -m e -M a.gemmell@physics.gla.ac.uk
with your e-mail address – this enables the batch system to send you an e-mail informing you of the completion (successful or otherwise) of your job.
Trainweights and Weights
Initial versions of the code used to have the weights for the ANN hard-coded into the runcards (General Parameter
FixWeight in the
FlatReader file). However, later versions of the code do not need this hardcoding – weights are calculated from values found in the input files themselves, and so
FixWeight has been set to 1 (this can be used to multiply certain samples' cross-sections by a given number, if desired). However, the formulae to obtain the weights are still quoted below, so that you can check the Computentp's work and make sure that it makes sense.
N.B. These weights are wrong for the ttjj (5212) sample. The input that was produced in v12 of athena was initially produced using
MC@NLO. This produces both positive (+1) and negative (-1) weighted events (an easy way to consider this is to consider the negatively weighted events as destructive events, that interfere with the positively weighted events, with the net result of decreasing the cross-section of the process). We considered all events equally for our calculation of the weights, simply considering the total number of events in the files. This problem will disappear when we switch to v15 inputs, where the ttjj samples have been produced using Alpgen.
The first weight to be considered is
TrainWeight – the scale factor we multiply each of the background events by so that they are in physically realistic ratios in relation to one another, while enforcing the requirement of the ANN training that we have equal numbers of signal and background events – used for the training of the ANN.
The calculation for the Trainweight is:
(Number of generated signal events / Number of generated events for that background) * (Cross-section for that background / Cross-section for all backgrounds combined)
The next weight is simply called Weight – this is the scale factor used to produce a physically realistic input for the ANN – now with the signal weighted as well. The formula used to find this is simply
(Number of events expected for your desired luminosity) / (Number of events present in input dataset)
Both of these numbers are calculated by Computentp, and can be checked in the file trees/NNInputs_120.root, where they are in branches labelled by 'TrainWeight' and 'weight'. However, it should be noted that while
TrainWeight as produced by Computentp is used by the ANN in the training sequence, the final results are produced independently Computentp – the ANN calculates the weight on its own.
Then, genemflat_batch_Complete2_SL5.sh creates a file called
FlatPlotterATLAStth${prefix}.txt . This file is used in the 'templating' phase of genemflat_batch_Complete2_SL5.sh and is based on the templates provided in teststeerFlatPlotterATLAStthSemileptonic-v15.txt.
And so via the
FlatPlotter file the
FlatReader files are included in the call:
runFlatPlotter \$steerPlotter ...
which produces a template for each of the signal and individual background samples.
User Setup
To set up the neural net,
-
check out the latest version from subversion (check what this is on trac
) using the command
svn co https://ppesvn.physics.gla.ac.uk/svn/atlas/NNFitter/tags/NNFitter-00-00-0X %BR%
-
ensure you know the ntuple_area variable to be passed in at run-time to genemflat_batch_Complete2_SL5.sh. This will be the directory where the input ntuples are stored (currently /data/atlas07/ahgemmell/NTuple-v15-30Aug for ntuples which have sensible states, events passing preselection and events failing preselection having my_failEvent == 3, 1 and 0 respectively – note that by their very nature, sensible states have also passed preselection).
-
The BASEBATCHDIR is now set automatically to the working directory when the script is executed.
Variables used by the Neural Net
The list of variables on which the neural net is to train is set in the shell script, under TMVAvarset.txt (this file is created when the script runs). At present, these variables are:
NN_BJetWeight_Jet1
NN_BJetWeight_Jet2
NN_BJetWeight_Jet3
NN_BJetWeight_Jet4
NN_BJetWeight_Jet5
NN_BJetWeight_Jet6
NN_BJet12_M
NN_BJet13_M
NN_BJet14_M
NN_BJet23_M
NN_BJet24_M
NN_BJet34_M
NN_BJet12_Pt
NN_BJet13_Pt
NN_BJet14_Pt
NN_BJet23_Pt
NN_BJet24_Pt
NN_BJet34_Pt
NN_State1_SumTopEt
NN_State2_SumTopEt
NN_State3_SumTopEt
NN_State1_DiffTopEta
NN_State2_DiffTopEta
NN_State3_DiffTopEta
NN_State1_DiffTopPhi
NN_State2_DiffTopPhi
NN_State3_DiffTopPhi
You also need to provide addresses to the Neural Net so that it can find the variables in the input trees. This is done inside
VariableTreeToNTPATLASttHSemiLeptonic-v15.txt
ListParameter EvInfoTree:1 1 NN_BJetWeight_Jet1:NN_BJetWeight_Jet1/NN_BJetWeight_Jet1
Currently all information is in the
EvInfoTree, which provides event level information. However, future work will involve trying to establish a
GlobalInfoTree, which contains information about the entire sample, such as cross-section - this will only need to be loaded once, and saves having to write the same information into the tree repeatedly, and subsequently reading it repeatedly.
Variable Weights in the Neural Net
To set up a neural net for the analysis of a particular kind of data it is necessary to train it with sample data; this process will adjust the "weights" on each variable that the neural net analyses in the ntuple, in order to optimise performance. These weights can then be viewed as a scatter plot in ROOT.
Specifying files as Signal/Background
The input datasets need to be specified in a number of peripheral files, so that the ANN can distinguish between signal and background files. Errors for each process also need to be specified - how this is done is detailed in that section. The relevant files for adding processes are atlastth_histlist_flat-v15.txt,
AtlasttHRealTitles.txt,
FlatAtlastthPhysicsProc1.txt and
FlatSysSetAtlastth1.txt. There are also some files that are produced through the action of genemflat_batch_Complete2_SL5.sh. At several points in these files, there are common structures for inputting data, relating to
ListParameter and
ColumnParameter:
ListParameter <tag> <onoff> <colon-separated-parameter-list>
<onoff> - specifies whether this parameter will be taken into consideration (1) or ignored (0) - generally this should be set to 1.
<tag> and <colon-seperated-parameter-list> - varies from process to process, will be explained for individual cases below. There can only be one instance of a <tag> active at any one time (i.e. you can write more than one version, but only one can be taken into consideration.
ColumnParameter <tag> <sequence> <keyword=doubleValue:keyword=doubleValue...>
The expression <tag>:<sequence> must be unique, e.g.
ColumnParameter File 0 OnOff=0:SorB=0:Process=Data
ColumnParameter File 1 OnOff=1:SorB=0:Process=Fake
where <tag> is the same, but <sequence> is different. The fact that the <sequence> carries meaning is specific to the implementation. Note that all of the values passed from
ColumnParameter will eventually be evaluated as Doubles - any variables where you pass a string (as for 'Process' above), this is not actually passed to the code - these code snippets are to make the code more easily readable by puny humans, who comprehend the meaning of strings more readily than Doubles.
atlastth_hislist_flat-v15.txt
This file provides a map for the ANN, giving it the output file names (and in which directory they are to be stored, relative to ${template_area} - set in genemflat) and the tree structure where the final result of the ANN will be stored in the output (in the example below, the output file is ${template_area}/11602-filter.root, and the result graph will be
FlatPlotter/NNScoreAny_0_0_0 0). The number to the left of the file name indicates which process it is - this is established using a file called TMVAsteer.txt, which is created through the running of genemflat_batch_Complete2_SL5.sh), and corresponds to the variable my_Eventtype in the input files (this can also be influenced in genemflat).
0 116102-filter.root FlatPlotter/NNScoreAny_0_0_0 0
The list of signal/background processes can be found in
AtlasttHRealTitles.txt (where the names are specified and associated with numbers). At present these are:
Process_0_0 TTjj:Semileptonic
Process_1_0 ttH:Semileptonic
Process_2_0 EWK:Semileptonic
Process_3_0 QCD:Semileptonic
genemflat_batch_Complete2_SL5.sh
genemflat creates the file TMVAsteer.txt, which sets a number of parameters for the running of the ANN - the constraints on the events, the precise Neural Net structure and so on - for establishing the input files, we are interested in only a couple of these parameters:
GeneralParameter string 1 FileString=my_Eventtype
Indicates the leaf in the input file which shows which process the event belongs to - this is the same number as we've specified as in atlastth_hislist_flat-v15.txt and
AtlasttHRealTitles.txt.
ColumnParameter File 1 OnOff=1:SorB=1:Process=tth
The number before the switches (
OnOff,
SorB, etc - in this case it is 1) corresponds to the number given in
AtlasttHRealTitles.txt. The other numbers are self-explanatory - they establish if that file is to be used, if it is signal or background (1=signal. 0=background) and the name of the process. In this instance, the Process name is just a comment for your own elucidation - it is not used itself in the code, so does not necessarily have to correspond to the process names as provided in
AtlasttHRealTitles.txt (though of course it is useful for them to be similar).
The other file that is produced by genemflat that specifies the input files for Computentp is steerComputentp.txt
# Specify the known metadata
ListParameter SignalProcessList 1 Alistair_tth
ListParameter Process:Alistair_tth 1 Filename:${ntuple_area}/ttH-v15.root:File:${mh}:IntLumi:1.0
This is just a list of the various input files, and we specify the integrated luminosity. The 'File' parameter is only used for book-keeping by Computentp, and does not have to correspond to the file numbers used in the ANN steering files (or to my_Eventtype), but for sanity's sake it is probably best to keep things consistent. We make an exception for the signal - we assign it the number ${mh} - so that we can keep track of things if we have different mass Higgs in our signals.
# Map of input file name to output file name: The ComputentpOutput will have a sed used to get the right mapping.
ListParameter InputOutputMapName:1 1 ${ntuple_area}/ttH-v15.root:${Computentpoutput}/tth_NNinput.root
The
InputOutputMapName is a list of integers - this doesn't have to bear any relevance to any numbers that have gone before - just give each output a unique number. This is followed by the mapping of input file names provided, to the output names that Computentp will produce.
This file contains various parameters:
ColumnParameter BackgroundList 0 tt0j=0
Here you specify once again the numbers assigned to the processes by my_Eventtype (for tt0j it equals zero), and list things as
BackgroundList or
SignalList. The number after 'BackgroundList' or 'SignalList' is unique for each process (to preserve the uniqueness of <tag>:<sequence>), but does not need to correspond to my_Eventtype. However, for completeness' sake within this file I have set it as such.
ColumnParameter PseudoDataList 0 tt0j=0
This is simply a restatement of the
BackgroundList - the same numbers in the same place. This list specifies the processes included in the pseudoexperiments, and therefore the signal process is not included in this list.
ListParameter ProcessLabels:1 1 tt0j:t#bar{t}0j
The number after
ProcessLabels again doesn't correspond to my_Eventtype - I have made it the same as the number after
BackgroundList/SignalList and
PseudoDataList. The important feature from this is that it tells the ANN what to label each of the various processes as in the results plots.
ColumnParameter UCSDPalette 0 tt0j=19
ColumnParameter PrimaryColorPalette 0 tt0j=0
These two parameters specify the colours used in the plotting for each of the processes (the numbers correspond to those in the Color Wheel of TColor). The numbers after the UCSDPalette and
PrimaryColorPalette are the same ones as have been used previously in this file. Whether the plotting uses the colours stated in UCSDPalette or
PrimaryColourPalette is determined in the file flatsteerStackNNAtlas.txt by setting the parameter:
GeneralParameter string 1 Palette=UCSDPalette
The final parameter to be set in
FlatAtlastthPhysicsProc1.txt is:
ColumnParameter ProcessOrder 0 tt0j=0
Once again, the number on its own (in this case 0) is the same as the other such instances in this file. The final number (zero in this case) is the order in which this process should be plotted - i.e. in this case, the tt0j sample will be plotted first in the output, with the other samples piled on top of it.
This file contains all the information on the errors that you pass to the ANN so that it can work out how the errors propogate to the final plots and answers, and so most of the details of this file will be covered in that section. The basic format of the file is:
ColumnParameter Combine:Lumi 0 OnOff=1:Low=-0.11:High=0.11:Channel=1:Process=TTjj
The <sequence> parameter (in this case '0') is there so that you can specify the parameters for a given error for multiple channels, without falling foul of the uniqueness requirement for <tag>:<sequence>. We have chosen it so that it equals my_Eventtype for that process. 'Channel' is present just in case you're considering multiple channels. We're only considering the one channel in this case (
SemiLeptonic). The final parameter (Process) is not actually used - the second parameter tells the ANN which errors are which, but this isn't very easily read by you, so feel free to add it in to help you keep track of the various errors! These final few parameters can be placed in any order, so long as they are separated by semicolons.
Passing preselection and sensible states
Many of the generated events in our samples will not pass the preselection cuts we would use in our final analysis. Sometimes to pass preselection requires some mistakes on the part of the reconstruction (e.g. tt + 0j), othertimes to fail preselection requires either the final state particles to be inherently unsuitable for our reconstruction, or to be mis-reconstructed. However, even if an event passes preselection it is possible that the events as reconstructed give a nonsensical final state - for example, the the light jets might not be able to be combined in such a way as to give a reasonable value of the W mass. Based on a few simple mass cuts, an event passing preselection can be determined to have a sensible state or not.
Currently, the type of event you are looking at is determined by looking at my_failEvent. States failing preselection have this equal to 0, passing preselection but not having a sensible final state equal 1 and passing preselection and having a sensible final state equal 3. These numbers are the basis of a number of bitwise tests - thus when setting your own my_failEvents, consider which bits in a binary string you want to represent various things, and then convert those to decimal.
Setting Systematic Uncertainties
The fitting code can take into account two different types of systematic uncertainty - rate and shape. The basic method to obtain both these uncertainties is that you should make your input samples for both your nominal sample, and for the two bounds of a given error (e.g. Initial State Radiation, ISR). Repeat this for all of the errors you wish to consider. The rate systematic uncertainty is simply how the number of events change that pass your preselection cuts etc. (you can only consider this, if you like). To obtain the shape uncertainty, you should pass each of the resulting datasets through the ANN (up to and including the templating, so that you have ANN results for both the nominal results, and as a result of varying each background). These ANN outputs can then be used to produce the rate uncertainties based on their integrals, before being normalised to the nominal cross-section so as to find the shape uncertainty - a measure of the percentage change in the bin-by-bin distribution for each error.
The fitting code is passed the relevant information about errors through the use of a number of files, but in the simplest case (when shape uncertainties are not being considered), there are only two:
FlatSysSetAtlastth1.txt and
SysNamesAtlastth1.txt. The basic call to the fitting code is in genemflat_batch_Complete2_SL5.sh:
sysfile=FlatSysSetAtlastth.txt
steerfile=FlatFitSteer.txt
mkdir -p templates/fit
rm -f templates/fit/out_${mh}.log
Fit ${basehistlistname} ${template_area}/ \$sysfile \$steerfile $mh > templates/fit/out_${mh}.log
The final call is rendered in the actual job file (e.g. run114) as
Fit /home/ahgemmell/NNFitter-00-00-09-Edited/NNTraining/atlastth_histlist_flat-v15.txt templates/tth120/ $sysfile $steerfile 120 > templates/fit/out_120.log
If you want to save time, (by not having to run templating for every error you wish to consider), you can instead only consider the rate uncertainties, and provide these as fractional changes to the rate, specified in
FlatSysSetAtlastth1.txt. Whether or not you consider shape uncertainties is controlled by a couple of parameters in the steering file
FlatFitSteer,txt, (which is created by the action of genemflat_batch_Complete2_SL5.sh)
GeneralParameter bool 1 UseShape=0
GeneralParameter bool 1 UseShapeMean=0
Setting
UseShape=1 means shape uncertainties will be taken into account for all the uncertainties that you provide the extra steering files and ANN scores for,
UseShapeMean=1 means that the ANN results for your various uncertainties will be used to produce the rate uncertainties based on their integrals, rather than on the numbers provided in
FlatSysSetAtlastth1.txt - using the relative sizes of the integrals of the AAN output as an estimator of the rate uncertainty can be useful if you don't want to be subject to statistical variations in the computation of your systematic uncertainties (if
UseShapeMean=0, the systematic rate uncertainty is calculated as a fractional change on the nominal rate). Considering shape uncertainties requires more steering files, and this will be detailed in later.
ColumnParameter Combine:Lumi 0 OnOff=1:Low=-0.11:High=0.11:Channel=1:Process=TTjj
The first parameter consists of two parts in this example: 'Combine' and 'Lumi'. The second part is the name of the uncertainty being considered. The first part 'Combine' (and the associated semicolon between them) is optional. It tells the ANN that the uncertainty thus labelled are independent of each other, and can be added in quadrature. 'OnOff' obviously tells the ANN to consider those uncertainty (1) or not (0). 'Low' and 'High' establish the relevant bounds of the uncertainty as fractions of the total (however, for the ANN these uncertainties are symmetrised, so to save time they are here assumed to be symmetric unless elsewhere stated) - note that these are not the uncertainties on the quantity, but rather the effect of that uncertainty on the rate of your process. Process is not actually read by the ANN, but is there to make the whole thing more human-friendly to read. The current errors, and their bounds are below. If no source for these error bounds is given, then they were the defaults found in the files from time immemorial (where as necessary I assumed that all tt + X errors were the same, as were all ttbb (QCD) errors, as in the original files the only samples considered were ttjj, ttbb(EWK), ttbb(QCD) and ttH - these errors probably originate from the CSC note). If you are only considering rate uncertainties, this is where the fitting code will find the relevant numbers.
ListParameter SysInfoToSysMap:1 1 Combine:LumiTrigLepID
The number in the <tag> after
SysInfoToSysMap is unique for each error (in this case it goes from one to eight). There is one entry per error considered, apart from the cases where the errors are combined in quadrature (as specified in
FlatSysSetAtlastth1.txt), where they are given one entry to share between them. The <colon-separated-parameter-list> provides a map between the name of the errors as considered by
FlatSysSetAtlastth1.txt (the errors combined in quadrature are lumped together under the name 'Combine'), and something more human-readable. The human-readable names are what will be written out by the fitting code (which identifies each error based on numbers, rather than the names in
FlatSysSetAtlastth1.txt) when it is producing its logfile. Obviously there is often not much change between the two names, apart form in the case of Combined errors.
Including shape uncertainties
For the sake of argument, we shall pretend to only be considering the one error overall. It is possible to consider rate errors independently of shape uncertainties (by setting
UseShapeMean=0 in
FlatFitSteer.txt) - this might be useful and quicker to run if a given error produces a large rate error, but the change to the shape of the ANN distribution is minimal (you can have a look at the ANN results yourself and make your own judgements). If you are not considering a given shape uncertainty but you are considering the rate uncertainty, all that needs to be done is to not produce the relevant steering files.
In this example, we will already have run three ANN templating steps - run1 (the nominal run), run2 (the results of taking the lower bound of the error) and run3 (the results of taking the higher bound of the error). You can then move into a new directory (e.g. run1_2) in which you want to perform the fitting, and at the very least set
UseShape=1 in
FlatFitSteer.txt (also perhaps setting
UseShapeMean=1). This requires some changes to the call to the fitting code and atlastth_histlist_flat-v15.txt so that the combination of ${template_area} and the filenames given in atlastth_histlist_flat-v15.txt still point toward the ANN template files you wish to consider - as shown in the two lines below (the first from genemflat establishing ${template_area}, the second from atlastth_histlist_flat-v15.txt establishing the filename for the ANN template):
template_area=templates/${process}${mh}
0 116102-filter.root FlatPlotter/NNScoreAny_0_0_0 0
could become:
template_area=${MAINDIR}
0 run1/templates/tth120/116102-filter.root FlatPlotter/NNScoreAny_0_0_0 0
This ensures that atlastth_histlist_flat-v15.txt will still point toward the ANN templates from the nominal run. You must now create additional steering files to point toward the high and low error ANN templates - their names are of the format:
"ShapePos_"+errorname+"_"+HistOutput
"ShapeNeg_"+errorname+"_"+HistOutput
where
HistOutput is atlastth_histlist_flat-v15.txt and errorname is the human-readable error name, as defined in
SysNamesAtlastth1.txt. You also need to change the ${basehistlistname} in the call to the fitting code so that it points directly at atlastth_histlist_flat-v15.txt, with no preceding directory structure - the code bases the names of the two extra shape steering files on this argument, and will not take into account any directories in the argument. (So that if ${basehistlistname} was directory/file.txt, the fitting code would look for the extra steering files with the name
ShapePos _ISR_directory/file.txt in the case of ISR being our error).
Filters
It is possible for the inputs to the ANN to have more events in than those that you want to pass to on for processing. We only want to train the ANN on those samples that would pass our preselection cuts and have 'sensible states'. The preselection cuts are easy enough to understand - they merely clean the sample - events which fail these are set to zero. However, it is possible for some events to have pass the cuts, but still not be anything like that which we would want (e.g. there is no way to reconstruct a top with a realistic mass). Seeing as it is likely that the backgrounds have more of these non-sensible states, to include them would be to give the ANN an unfair advantage in determining signal from background. But we don't want to simply set them equal to zero, as at other times we are going to be interested in how events passing preselection can fail to have sensible states, and so will want to examine them further. We therefore have filters so that Computentp and the ANN only look at events of our choosing. These filters take the place of various bitwise tests in genemflat_batch_Complete2.sh (not currently used, as explained below) and
TreeSpecATLAStth.txt.
In
TreeSpecATLAStth.txt the filters control what is used for the templating, and Computentp:
ListParameter SpecifyVariable:Higgs:cutMask 1 Type:int:Default:3
ListParameter SpecifyVariable:Higgs:invertWord 1 Type:int:Default:0
InvertWord is used to invert the relevant bits (in this case no bits are inverted) before the cut from cutMask is applied. The cutMask will exclude from templating those events where the matching bits are equal to zero AFTER the inversion. So here, with no inversion applied, those events with my_failEvent == 3 will be used for templating.
The constraint
GeneralParameter string 1 Constraint=(my_failEvent&3)==3
in genemflat_batch_Complete2.sh controls the events used in the training, using a bitwise comparison. If the constraint is true (i.e. the first two bits are set, and not equal to zero), then the event is used for training. This filter is not used currently, as training of the net takes place based on the Computentp output - this Computentp output only contains sensible states (as specified in the
TreeSpecATLAStth.txt file's filter). If further filtering is required, then care must be taken to ensure that my_failEvent (or whatever you wish to base your filter on) is specified in the
VariableTreeToNTP file, so that Computentp will copy it into its output.
If USEHILOSB is set to 1 then && must be appended to cut criteria, e.g.
GeneralParameter string 1 Constraint=(my_failEvent&65536)==0&&. This is because USEHILOSB adds more constraints.
Running
To run the script, first log into the batch system (ppepbs).
The genemflat_batch_Complete2_SL5.sh script can be executed with the command:
./genemflat_batch_Complete2_SL5.sh 12 480 1.0 tth 120 120 1 1 /data/atlas07/ahgemmell/NTuple-v15-30Aug
These options denote:
-
12 is the run number
-
480 is the jobstart - this is a potentially redundant parameter to do with the PBS queue.
-
1.0 is the luminosity will be normalised to (in fb^-1).
-
tth is the process type - aim to develop this to incorporate other processes, e.g. lbb
-
120 is the min. Higgs mass
-
120 is the max. Higgs mass
-
1 indicates that training is to take place
-
1 indicates that templating is to take place
-
/data/atlas07/ahgemmell/NTuple-v15-30Aug is the directory where the input ntuples are located (having my_failEvent bits set for ( 65536 for >0 sensible states) and ( 131072 for 4 tight b-tagged jets)
Note that it is possible to omit the training in subsequent (templating) runs, if it has previously been done (set 3rd from last argument to 0). This shortens the run time significantly.
Once the job has been completed you will receive an email summarising the outcome.
Running:
-
Creates a run12 subdirectory in working directory and makes it the working directory
-
Creates TMVAsteer.txt - writes fitting parameters to it
-
NN structure is set ( H6AONN5MEMLP MLP 1 H:!V:NCycles=1000:HiddenLayers=N+1,N:RandomSeed=9876543). This line sets up two hidden layers with N+1 and N neurons respectively (where N is the number of input variables).
-
Training cycles (1000) and hidden layers - N+1?
-
4 text steer files are copied into the run directory for templating
-
2 text steer files are copied into the run directory for stacking plots.
-
2 lines of text are appended to a temporary copy of flatsteerStackNNAtlas.txt: GeneralParameter string 1 HWW=tth-TMVA
GeneralParameter double 1 <nop>IntLumi=${lumi}
Contents of training file:
-
jetmin/jetmax - These seem to be redundant. Commented out, effective as of v.3
-
zmin/zmax (1/2) - what is their function?
-
weighting = <nop>TrainWeight - is this redundant?
-
TMVAvarset.txt - input variable set
Where the output is stored
-
trees/NNInputs_120.root
The output from Computentp - it is a copy of all of the input datasets, with the addition of the variables TrainWeight and weight.
-
computentp_output/
The same as for trees/NNInputs_120.root, but now each individual input dataset has its own appropriately named file.
-
weights/TMVTEST2_120_3.root_H6AONN5MEMLP.weights.txt
Contains the weights of the Neural Net - results of the training. Starts with info about the run: Date, time, etc. followed by various parameters you've set. Then the variables you're considering in the net, with their range. Finally comes the Neural Net structure itself - weights etc.
-
TMVA2_120_3.root
-
InputVariable _NoTransform/my_----_B_NoTransform
The distribution of this variable for the Background (or signal for S). N.B. Has N_sig / 2 entries (other half used for testing, not training). Check this to make sure there are no obvious problems with the input - missing data points, random spikes, etc.
-
Method_MLP/H6AONN5MEMLP/estimatorHistTrain
A measure after each iteration of how much tweaking is required to get S=1, B=0. Should settle after a while to a stable number
-
Method_MLP/H6AONN5MEMLP/estimatorHistTest
Same, but for the test sample - should look broadly similar. Otherwise we're over/undertraining.
-
Method_MLP/H6AONN5MEMLP/MVA_H6AONN5MEMLP_B
The final Neural Net Score for the Background (or signal for S) (Test result - not the final result).
-
Method_MLP/H6AONN5MEMLP/MVA_H6AONN5MEMLP_B_high
Rick's not sure.
-
Method_MLP/H6AONN5MEMLP/MVA_H6AONN5MEMLP_rejBvsS
Comparing Bkg rejection to Signal Efficiency. Ideally want rej=1 with eff=1 (i.e. reject all bkg, accept all signal).
-
train2.log
A text summary of the training process. Mentions files used (how many files, numbers of events, if they signal/background, etc). Will refer to file 0, 1... - refer to Atlas ttHRealTitles.txt to work out which file is which. Use this to make sure the right numbers of events are getting though your filters.
-
templates/out/FlatPlotter${prefix}.out
The output from the screen from the templating stage, where the Neural Net scores are actually produced, detailing errors etc.
-
templates/ttH120/
Files showing the final results of the ANN. NNScoreAny _0_0_0 is of particular interest - this is the ANN output for that file, weighted to represent real data (i.e. if you simply add all of these together, you should get a realistic 'Signal and Background' result, which is shown in:
-
stacked/Plots_120_TMVA_Lum_1.0/FlatStack_1.eps
stacked/Plots_120_TMVA_Lum_1.0/FlatStack_2.eps
The final ANN scores of the signal and background, scaled as to real data
-
stacked/out1_120.log
stacked/out2_120.log
Output from making the stacked (combined) plots.
-
templates/fit/out_120.log
This is the paydirt - the output to screen from the fitting stage. Look at the end of this file, and you shall see the exclusions generated by all the pseudoexperiments performed on the ANN output, in a little table, also giving +/- 1/2 sigma results.
-
drivetestFlatFitAtlastth.rootUnscaledTemplates.root.
The distribution of the exclusions. This is plotted only if
GeneralParameter bool 1 PlotLikelihood =1
in teststeerFlatReaderATLAStthSemileptonic.txt. The range and number of bins in this plot can be controlled by editing the following switches: GeneralParameter int 1 LikeliPseudoExpNBin=400
GeneralParameter double 1 LikeliPseudoExpMin=0.
GeneralParameter double 1 LikeliPseudoExpMax=10.
-
tr[run number].o[PBS job number]
This output file is written as soon as the job stops, and contains a summary of the full run - useful for looking for error messages.
-
TMVAPerf_120_Run[run number]Job2.html
A useful little html page that one of Rick's scripts creates, showing a number of useful plots - the signal and background Net scores, distributions of input variables and their correlations, and so on.
Limitations
-
It must also be run on a PBS machine because of the structure of the genemflat_batch_Complete2_SL5.sh file (i.e. PBS commands).
-
The file <nop>teststeerFlatPlotterATLAStthSemileptonic.txt appears to contain an invalid range for the pseudorapidity (max. value = pi)
-
If USEHILOSB is set to 1 then && must be appended to cut criteria, e.g. GeneralParameter string 1 Constraint=(my_failEvent&65536)==0&&
-
It would be desirable to adapt the code to be able to process different signals, e.g. lbb.
Diagnostic Run
A diagnostic run may be carried out by setting Debug=1 and NEvent=99 in teststeerFlatReaderATLAStthSemileptonic.txt. It is also advisable to cut the run time down by setting the number of training cycles to a low number (e.g. 20) in genemflat_batch_Complete2_SL5.sh - this appears as NCycles in the TMVAsteer.txt part of the file.
TMVA Training Plots
There is a macro in
the latest NNFitter version
which will plot the contents of run12/TMVA2_120_3.root showing the responsiveness of the neural net as a function of the number of training cycles, in order to gauge the optimal number of cycles to use (i.e. avoid the dangers of under- or over-training). The macro must be run in a directory where a neural net run has already been carried out.
This will create two .eps output files, one showing the success of signal/background fitting, and the other displaying the sensitivity of the neural net to the number of training cycles used, allowing the speed of convergence to be gauged.
Running analysis & making ntuples
This is the procedure used to remake the ATLAS ntuples, using code from CERN Subversion repositories. These ntuples were then used as input for the neural net.
cd ~
mkdir tth_analysis_making_ntuples_v.13
cd tth_analysis_making_ntuples_v.13
export SVNROOT=svn+ssh://kirby@svn.cern.ch/reps/atlasoff
svn co $SVNROOT/PhysicsAnalysis/HiggsPhys/HiggsAssocTop/HiggsAssocWithTopToBBbar/tags/HiggsAssocWithTopToBBbar-00-00-00-13 <nop>PhysicsAnalysis/HiggsPhys/HiggsAssocTop/HiggsAssocWithTopToBBbar
cd <nop>PhysicsAnalysis/HiggsPhys/HiggsAssocTop/HiggsAssocWithTopToBBbar/NtupleAnalysis/
make
In a new terminal window: cd /data/atlas07
mkdir gkirby
cd gkirby
mkdir ntuples_sensstatecutword
The script files in the <nop>NtupleAnalysis directory were then altered to output to this new directory. This is the output line for the signal ( tthhbbOptions86581430018061-nn.txt)
OUTPUT /data/atlas07/gkirby/ntuples_sensstatecutword/86581430018061-nn.root
Code changes: The tthhbbClass.cxx file was edited to include a new error code: the following lines were added to it to allow us to exclude events that we do not wish the NN to train/test with.
if (<nop>SensibleStates.size()==0) {
m_failEvent+=65536;
}
Then the make command was used again in this directory. The tthhbb executable was run with each of the input ("Options") text files to prepare the ntuples.
Another change was also required; the tthhbbClass.cxx file was also edited to include a fail code to allow for events not having '4 tight b-tagged jets' since this was one of the criteria used in the cut-based preselection. The following code was added to tthhbbClass.cxx:
if (<nop>BJets.size()<4) {
m_failEvent+=131072;
}
This was done so that the Signal to Background ratio could be increased in order for the fit to finish and provide sensible results. Requiring 4 tight b-Jets removes proportionally much more of the ttjj background than the other samples because there are fewer b-jets. The reason the S/B ratio was so low was that a problem was found which meant that the original ratio used in the 'fix_weight' variable in genemflat_batch_Complete2_SL5.sh (up to version -06) was in fact incorrect, and when the correct weights were calculated the S/B ratio was very low indeed.
The Neural Net was then configured to exclude events where: (m_failEvent & 196608)==1
with 196608=131072+65536.
Debugging the code
To debug the code, two things need to be done - first, all the debug switches need to be turned on, and then you need to restrict the number of events to ~10 (for a Computentp run this will still manage to generate a 2 GB log file!). All of these switches are found in teststeerFlatReaderATLAStthSemileptonic.txt (the progenitor for all
FlatReader files). The debug switches are:
GeneralParameter bool 1 Debug=0
GeneralParameter bool 1 DebugGlobalInfo=0
GeneralParameter bool 1 DebugEvInfo=0
GeneralParameter int 1 ReportInterval=100
All the debug switches can be set to one (I'm not sure of the exact effect of each individual switch) - the report interval is probably best left at 100. To restrict the events you use
#
# Loop Control
#
GeneralParameter int 1 NEvent=999999
GeneralParameter int 0 FirstEvent=1
GeneralParameter int 0 LastEvent=10
The easiest switch is to set NEvent=10 - however, if desired you can run over a specified range, by switching of the NEvent switch (changing it to int 0 NEvent) and switching on the other two switches, using them to specify the events you wish to run over.
Various other switches of interest
In
FlatReader:
GeneralParameter int 1 LikeliPseudoExpMin=0.
GeneralParameter int 1 LikeliPseudoExpMax=10.
GeneralParameter int 1 LikeliPseudoExpNBin=400
To enable you to specify the range and number of bins in the histogram showing the distribution of the pseudoexperiment exclusions. (Found in drivetestFlatFitAtlastth.rootUnscaledTemplates.root)
In both
FlatReader and
FlatPlotter:
GeneralParameter bool 1
LoadGlobalOnEachEvent=1
This needs to be set to one if your input datasets have non-sensible states. Otherwise, global variables are loaded on one occasion from the first entry in the file, and kept. If this entry is not sensible, or for some other reason has an unreasonable answer for this global value, a problem will develop. With this switch on, the values will be loaded each and every time – obviously this slows the code down – if the global values are safely stored in every entry, it might be best to set this to false.
TMVAsteer.txt (genemflat_batch_Complete2_SL5.sh)
H6AONN5MEMLP MLP 1 !H:!V:NCycles=1000:HiddenLayers=N+1,N:RandomSeed=9876543
If the phrase 'H6AONN5MEMLP' is changed, then this change must also be propogated to the webpage plotter (e-mail from Rick 1 Mar 2011)