Software Procedure

During initialisation the routine CLN_CLASSIFIER_INI reads the control titles bank TCLN 1 and, if it exists, the TCLN 2 bank describing the network (weights, biases, etc.) In addition, the banks to be used for storing the network input and output are defined (CLNP and CLNO). Finally, an HBOOK output stream may be created, which may be used for histogram output . The user may enable histogram outputting via word 7 of the TCLN 1 bank, and should specify an output filename in her command file by ``FILE CLN 1 filename''. Not all parameter types produce histogram output.

In the execution phase CLN_CLASSIFIER calls CLN_CLASSIFY_EVENT for each event. This second routine first checks that the number of hits lies in the range set by words 1 and 2 of the TCLN 1 bank. It will then look for a fitted vertex and direction if either are required in the calculation of the parameters (the type of fit being set by words 3 and 4 of the TCLN 1 bank). Given a fitted vertex, the pmt hits are copied into a private array throwing away any which do not pass the residual cut

$$\left\vert\vert\underline{r}_{fit}-\underline{r}_{i}\vert-v(t_{fit}-t_{i}) \right\vert < R_{cut}\\$$ (18.1)

where $v$ is the speed of light in water ($R_{cut}$ is set by word 5 in TCLN 1). The surviving hits in the pmt array are then available for use in calculating the parameters. Some parameter types, such as ACF (Angular Correlation Function) and ITR (Intime Ratio), do not use this array, but instead use all tubes in the event, or apply their own cuts. The calculation of these parameters is thus unaffected by the value of $R_{cut}$.

The actual parameters to be extracted are specified by the user in the TCLN 1 titles bank. They may include single parameters such as nhit or the angle between the fitted direction and the sun. There are also sets of parameters of a particular type such as the harmonic parameters (these encode the angular distribution of the hit pattern). Some parameters need to be turned on in sets (for example, the user should always specify either 0 or 14 ACF parameters, and the code will not accept any other value.) In addition, the user may define new parameters not included in the standard set (see section 18.2.4 for how to do this). The routine CLN_GET_PARAM makes calls to separate modules (having a generic name CLN_PARAM_xxx) which calculate a particular type of parameter.

Finally, the user determines where the parameters are stored and whether the network fitter is to be applied using the command:

$cln_network $update_fitter_banks

or

$off

or

$on

• Selecting $update_fitter_banks results in the classifier searching the FT (Fitter Dispatch bank) for the current event and, for each fitter found, creating an FTXC (Fitter Classifier Parameters bank). The fitter's own position and direction are used as input to the parameter definition; the fitters specified in the TCLN 1 bank are ignored. The format of FTXC is similar in content to the CLNP (Network Classifier Parameters bank) but is fixed format making it easier to access.

• Selecting $off results in a single CLNP being produced using the position and direction from fitters as specified in the TCLN 1 bank.

• Selecting $on produces a CLNP as in $off but then the routine CLN_CALC_NET_OUTPUT will feed the input parameters through the network defined by the weights and thresholds in the TCLN 2 bank. The network is currently limited to having only 1 hidden layer and a unit activation function of the form

  
  

  $$Y = [ 1 + \exp (- \sum_{i} w_{i}X_{i} - \theta )]^{-1}$$ (18.2)

  
  

  where $w_{i}$ are the weights linking the unit to the outputs $X_{i}$ of the previous layer.