I thought it worthwhile to enumerate some of the different algorithms that I have tested.  Code for all algorithms is found in $CVSROOT/offline/users/sgliske/StRoot/StEEmcPool/StEEmcHitMaker.

Usual Method + New Clustering

The following ESMD strip clustering methods were tested, using all other steps of the IU pi0 reconstruction algo the same.  Note: most of the following methods were scrapped due to the fact they worked poorly.  Methods that were tested include:

• The IU method: a seed strip type of method
• k-means++, i.e. the usual k-means plus a fancy initialization procedure and cost function from D. Aurthur (http://www.stanford.edu/~darthur/kmpp.zip)
• Gaussian Mixture Model, plus various criterion, such as the AIC and BIC and variants, for determining the optimal number of clusters
• Morhac's method, i.e. using the TSpectrum peak finder in root
• A simplified seed strip method (called the "Naive" method)
• A more advanced seed strip method (called ANLS) which merges clusters based on the energy density between peaks.
• Tukey-Smoother (i.e. TH1::SmoothArray), and then reading off the peaks
• Tukey-Smoother initializing a Gaussian mixture model
• Tukey-Smoother plus some other features, i.e. Tukey-Smoother Plus (TSP)

In general, it is very difficult to have an algorithm that both does not falsely split single clusters and also does not falsely merge multiple clusters into a single cluster, and that can do this over a range of pT.

ANLS did OK in most things.  It still have some trouble with split clusters and the asymmetry in the photon pairs was to small, yielding too high of an invariant mass.

The TSP seems to be the most optimal method considered thus far.

Iterative Methods

An iterative method was described at the Nov. collaboration meeting.  Several different algorithms were considered for deciding when to split a hit into a pair of hits.  Method considered include:

• An Eigen-value method, similar to that used by S. Heppleman
• Morhac's peak finder method, i.e. the TSpectrum peak finder
• Comparing the quartile positions with that which one would expect from a Gaussian distribution
• Cutting on the interquartile distance
• Using the ANLS strip cluster algorithm

It may also be interesting to use the TSP strip cluster method and see if the iterative method has advantages over the linear method.  This has not been tried as of date (3 Jan, 2012).