titles_tcln

                              Titles:  TCLN
                              =============
 
Contact:  M. Thorman, Oxford.
 
Revision History:-
================
 
  2.09  M.Thorman  First version.
  3.00  M.Thorman  Added KTCLN_LEG_PAR word to TCLN 1.
        X.Chen     Added KTCLN_THETA_IJ word to TCLN 1.
  4.01  N.McCauley Added KTCLN_NECK_LIKE_PAR word to TCLN 1. 
  4.02  B. Heintzelman Added KTCLN_ANG_CORR_HISTOS word to TCLN 1.
                       KTCLN_ANG_CORR_HISTOS controls histogram ID offset.
        S. Oser    Added ITR (InTime Ratio) parameterisation.
        N. West    Fix type (SOL_DIR_PAR and HAR_PAR were interchanged).
                   Add Network on/off mode: Update fitters.
        B. Heintzelman ANG_CORR now 17 parameters.

 
Titles Files
------------
 
classifier.dat
 
Verification
------------
 
 
Description
-----------
 
There are actually two banks for the network classifier. TCLN 1 is a  general
control bank used to specify which parameters are to be extracted from each
event and also which fitters to use. TCLN 2 holds a complete description of a
feed-forward neural network with any number of input, hidden and output nodes
(but only one hidden layer). The activation function used at each node is
 
               Act = 1 / ( 1 + exp-(w.X + bias) )
 
Where (w.X) is the weighted sum of the outputs of all nodes from the previous
layer.
 
Refer to the symbol table for the relevant symbolic commands.
 
Data Words
---------
 
   TCLN 1
   ======
 
        Fixed length Database Header.  See titles_dbhdr.html
 
   KTCLN_
 
1   I   MIN_NHIT     Minimum number of pmt hits for classification
2   I   MAX_NHIT     Maximum number of pmt hits for classification
3   I   VTX_TYPE     Type of fitted vertex to use    (-1..9):
4   I   DIR_TYPE     Type of fitted direction to use (-1..9):
                       -1 = MCO
                        0 = not required
                        1 = FT1
                        2 = FT2         (+ve values correspond to the
                        3 = FT3          structural link from the FT bank
                        4 = FTE          for the desired fitter.)
                        5 = FTT
                        6 = FTQ
                        7 = FTG
                        8 = FTM
                        9 = FTP
5   F   RES_CUT      Residual cut (cm) - if a pmt hit has a residual exceeding
                        this value, it is discarded and not used when
                        calculating the  parameters.
6   I   USE_NET      Network on/off
                        =0  if just generating the input parameters
                        =1  if network output also required (A network must be
                               defined in TCLN 2 for this case)
                        =-1 if FTXC (Fitter Classifier) banks are to be produced
                               instead of CLNP/CLNO.
7   I   ANG_CORR_HISTOS  
                     Angular Corr Fn histogram control:-
                        =0   none output;
                        >0   histogram numbering offset i.e. all histograms
                             have this offset added
8   I   ITR_LOWER_WINDOW ITR lower time window
9   I   ITR_UPPER_WINDOW ITR upper time window

10-14                Spare

 
        Selecting the network input parameters:-
 
15  I   NHIT_PAR     NHIT parameter (switch)
16  I   SOL_DIR_PAR  Cosine of angle to sun (switch)
17  I   HAR_PAR      Number of harmonic (beta) parameters
18  I   RES_HAR_PAR  Number of residual-harmonic (gamma) parameters
19  I   RES_PAR      Number of residual (mu) parameters
20  I   LEG_PAR      Number of Legendre polynomial parameters. (See note 1.)
21  I   THETA_IJ_PAR Theta_ij parameter (switch)
22  I   NECK_LIKE_PAR Neck Likelihood Parameter (switch) 
23  I   ANG_CORR_PAR Heintzelman angular correlation parameters 
                     (must be either 0 or 17)
24  I   ITR_PAR      Intime Ratio (ITR) parameter

Notes

1. The Legendre polynomial parameters actually give the same information as the harmonic parameters (there is a simple linear relationship between the two). They are, however, much simpler to interpret, just being the mean value of P_l(cos{theta_ij}) for all pairs of hits i and j without any scale-adjustment or normalization factors. (P_l is the Legendre polynomial of degree l and theta_ij is the angle between hits i and j with respect to the fitted vertex.)

 
 
    TCLN 2
    ======
 
        Fixed length Database Header.  See titles_dbhdr.html
 
   KTCLN_
 
1   I   NUM_INP     Number of input nodes (An error will be generated if this
                        does not match the requested number of parameters.)
2   I   NUM_HID     Number of hidden nodes
3   I   NUM_OUT     Number of output nodes
4-7                 Reserved
 
****  1st hidden node:
 
8   F   WEIGHTS     Bias
9   F               Link to input 1
10  F                             2
11  F                             3
12  F                             .
 .                                .
 
****  2nd hidden node:
 
Ni+1 +8   F         Bias
      9   F         Link to input 1
      10  F                       2
      11  F                       3
      12  F                       .
                                  .
 
	...similarly for all remaining hidden nodes
 
****  1st output node:
 
Nh(Ni+1) +8   F     Bias
          9   F     Link to hidden node 1
          10  F                         2
          11  F                         3
          12  F                         .
                                        .
        ...similarly for remaining output nodes
 
Here, Ni and Nh indicate the number of input and hidden nodes.
 
Note: Even if there is no link between layers for some nodes, a weight should
still appear here, and be set to zero.