Home » Blogs » bouchet's blog

Neural Network to predict SST hit (run 14 preproduction)

Updated on Sun, 2016-01-31 15:21. Originally created by bouchet on 2016-01-31 14:26.
  • using R
  • Neural Network (MLP, activation function = sigmoid, 3 hidden neurons)
  • data : day 160, divided in 3 Pt bins
  • features : vx,vy,vx (vertex of the track), px,py,pz (momentum), # of TPC hits, pseudo-rapidity, isPxlInner (0,1), isPxlOuter(0,1),isIst
plots are the NN layout and the residuals (Actual - Predicted)

1) 0 < Pt < 1
training : 25000 tracks
testing : 5000 tracks



2) 1 < Pt < 2
training : 3000 tracks
testing : 500 tracks


3) 2 < Pt < 3
training : 250 tracks
testing : 60 tracks


Comments, Results
The first Pt bin has a pretty large residuals as seen by the NN output

> nn1
Call: neuralnet(formula = sst ~ vx + vy + vz + px + py + pz + nHits +     eta + pxlInner + pxlOuter + ist, data = training, hidden = 3,     threshold = 0.01, lifesign = "minimal", linear.output = FALSE)
1 repetition was calculated.

        Error Reached Threshold Steps
1 1546.127789    0.008942761227 16219
But the 2 next bins have decent residuals, as seen also by the NN outputs :

> nn2
Call: neuralnet(formula = sst ~ vx + vy + vz + px + py + pz + nHits +     eta + pxlInner + pxlOuter + ist, data = training2, hidden = 3,     threshold = 0.01, lifesign = "minimal", linear.output = FALSE)
1 repetition was calculated.

        Error Reached Threshold Steps
1 96.63010408    0.009238088031  8963


> nn3
Call: neuralnet(formula = sst ~ vx + vy + vz + px + py + pz + nHits +     eta + pxlInner + pxlOuter + ist, data = training3, hidden = 3,     threshold = 0.01, lifesign = "minimal", linear.output = FALSE)
1 repetition was calculated.

        Error Reached Threshold Steps
1 2.503906866    0.009956694217 25495
R console is here

4) Quick test with Decision Tree for 1 < Pt < 2 bin
(rpart from the caret package)
Note : pxlOuter, pxlInner, ist have to be treated as categorical features (0 or 1)

> m4<-train(sst ~ as.numeric(vx) + as.numeric(vy) + as.numeric(vz) + as.numeric(px) + as.numeric(py) + as.numeric(pz) + as.numeric(nHits) + as.numeric(eta) + as.factor(ist),data=training2,method="rpart")
> m4
CART 


3000 samples
  11 predictor
   2 classes: '0', '1' 

No pre-processing
Resampling: Bootstrapped (25 reps) 
Summary of sample sizes: 3000, 3000, 3000, 3000, 3000, 3000, ... 
Resampling results across tuning parameters:

  cp              Accuracy      Kappa         Accuracy SD    Kappa SD     

  0.007523939808  0.7891367251  0.4278404062  0.01325860196  0.03788879076
  0.041039671683  0.7846494110  0.4254553030  0.01222009237  0.06024589731
  0.063383492932  0.7672447407  0.3242443199  0.01672489871  0.19504303949

Accuracy was used to select the optimal model using  the largest value.

The final value used for the model was cp = 0.007523939808.



 





»