Better regressor for Disp

[mahnen]

Hi guys,

I wrote a little something about the current status of the Disp-parameter. In particular it would be quite important to get this Parameter estimated better.

disp_rf.pdf

[mahnen]

small summary from the pdf:

• I to trained a disp RF and applied it to the mc with facttools
• Target for training: 'Distance' in MC of gammas equivalent to Kai's approach, use sklearn and then export to .pmml
• I used the 'apply model' processor in the jpmml branch.

Results:

• PSF got a bit better (~16%) → therefore bg now down to: ~ 75% of its value before, smaller tails. (probably much improvable with better disp-models)
• Problem of loss of gamma rays at highes energies after theta-cut: gone!

My opinion on it:

• The .pmml – apply model processor in facttools is super-useful. We should definitely use it!
• You can check out the DispRF-xmls I used in the jpmml branch of facttools. The regressor itself I uploaded to the Phido:

/fhgfs/groups/app/fact/simulated/DispRegressor_v1.pmml

• Imho this is the last issue in facttools which, when solved, would make this the superior performing analysis in comparison to MARS.

[maxnoe]

Hey,

Can you give more details on the plots you showed? I do not really understand what the different colors mean. Can you maybe add a plot where you just show the comparison for the old disp method and your random forrest?

Also a correlation plot between the two methods might be interesting.

[mahnen]

1) correlation plot between the two methods

2) comparison for the old disp method and your random forrest

[mahnen]

I made some more plots. In theory, Disp = Distance for gamma-rays. The following plots are Disp vs Distance for the two methods. It looks to me rather like the current parametrization of Disp is a random number...

[mahnen]

The last comparison is the size over theta^2 distribution. It shows that we can particularly gain something at higher sizes

[ftemme]

Really nice, but I have some comments on that:

• I don't think that Disp = Distance, cause Distance doesn't take the angle to the source (alpha) into account.
  • Maybe it is enough to use Disp = Distance / cos(alpha)
• In the MCs, Distance (and alpha) are already based on the reconstruction of the event (so extraction image cleaning ...) so I have difficulties to think about what is really the 'mc-truth' for disp.
• I have difficulties to think about the sign for the source reconstruction (so should disp be added in one or the other direction)
• Just as an information for all of you: The currently used disp parameterization is a very easy one only using width and length. Nice to see, that is has performance issues at high energies, so there is quite room for improvement left.

I would propose to have a brainstorming telcon (or skype con) next week (11.4. - 15.4.) to discuss this. But I agree with you, this is one of the main remaining issue FACT-Tools has to handle and your application of the RF regression is quite impressive.

[mahnen]

I have looked up the definition of Dist in Sebastian Commichaus old PhD thesis. He there writes:

The image parameters DIST and ALPHA are so-colled source-dependent image parameters...

The drawing in my mind goes as follows:

Also, the current implementation of the alpha (by Kai?) looks source dependent:

`

    //TODO: this might throw an NPE for source[1]

    double auxiliary_angle  = Math.atan( (source[1] - dist.getCenterY() )/(source[0] - dist.getCenterX()) );    

    //auxiliary_angle         = auxiliary_angle / Math.PI * 180;

    alpha                  =  (dist.getAngle() - auxiliary_angle);

`

and also our implementation of Dist is source dependent, as Kai writes:

/**

• Quite simply the distance between the CoG of the shower and the calculated source position.
• @author kaibrugge /

So after all it looks to me that Dist for Gamma-MC is the MC-truth for Disp.

Sebastian C. further continues about the sign of the Disp parameter (M1 - single telescope!):

Because of the head-tail ambiguity in the 2nd moment analysis, the Disp method provides two possible solutions for the location of the source along the mayor axis of the Hillas ellipse. A possibility to overcome the head-tail ambiguity is to make use of the asymmetry parameter M3LONG. ... The achieved ratio of correct head-tail assignment is in the order of 70%, which strongly depends on the considered energy range.

The MC-truth (DIST) is really, like you mentioned, cleaning dependent etc..., which unfortunately breaks the streams-flow... you have to process it first and then build the Disp-RF and then run again.

Telcon next week - no problem.

[maxnoe]

Moved to our python classifier tools and works quite nicely