Addendum to 31Jan05 presentation

02 Feb 2005

This is a small addendum to my talk to the Ana Group. Bill Atwood pointed out that the evaluation of the decision tree was flawed;  it used the same data to evaluate the separation as was used to create the tree. This is bad since the procedure is sensitive to the training set.

I responded that I had limited nodes to have at least 100 events, but did not know the answer to his question. So here are some results. I train on only the even events in the all_gamma data set, then run the evaluation on three data sets:

• the even events, which should show the bias
• the odd events, which should be unbiased
• uncorrelated events (choosing the next entry in the tuple) which should show no correlation

The plot follows:

Conclusions:

• The testing sample, using events not used for training, is worse, but not by much, since the end nodes were limited to 100 events
• Applying the good gamma prediction to the wrong events has no effect: the result is really uncorrelated

--Toby Burnett