The atlantics

[damonge]

Selection method developed across the Atlantic ocean by a bunch of us.

[EiffL]

Thanks for entering the challenge :-D
Whenever you are ready and have a few results, feel free to document your approach in this PR, to help other people get a sense of what the competition is ;-)

[damonge]

Yep, will do @EiffL ! I saw an unofficial statement from @slosar on slack that the deadline for this could be pushed to Sept ~15th. Is that the case?

[EiffL]

Undecided I think ^^' I'm gonna bring that issue up on slack.

[damonge]

Summary of the method: We try to find an optimal set of bins in three steps

1. Create a SOM
2. Bunch up different SOM cells into a set of O(100) groups. The groups should be chosen wisely. Two options are grouping them through a kd-tree descent by Euclidean distance in color space or in "N(z)-moment" space.
3. Bunch those groups up again into a number Nbin of tomographic bins in a way that maximizes the total S/N. Given a set of Nbin-1 redshifts defining the bins, the groups are initially assigned to bins based on the bin their mean redshift falls into. Then, groups can be put into a trash bin if their overlap with their bin is not large enough, or if their overlap with far-away bins is too large. We have thus Nbin+1 free continuous parameters: the Nbin-1 edges, the minimum fractional overlap within the bin, and the maximum fractional overlap in far-away bins. We vary these parameters to maximize S/N.

Note that step 2 is only there so we can deal with a reasonably small set of groups where we can precompute all auto- and cross-correlations, which allows us to speed up the calculation of S/N significantly.

[joezuntz]

Hi all. To make ComplexSOM work I had to make this change:

+++ b/tomo_challenge/classifiers/ComplexSOM.py
@@ -402,7 +402,7 @@ class ComplexSOM(Tomographer):
         cumul_f = interp1d(z_cen, cumulative_fraction, bounds_error=False,
                            fill_value=(0, 1))
         edges_0 = np.array([brentq(lambda z : cumul_f(z) - q, 0, 2)
-                            for q in (np.arange(o.n_bins-1)+1.)/o.n_bins])
+                            for q in (np.arange(n_bin-1)+1.)/n_bin])
         
         # Finally, let's write the function to minimize and optimize for a 4-bin case.
         def minus_sn(edges, calc):

otherwise it crashed with NameError: name 'o' is not defined

can you confirm that this is what was intended?

[damonge]

Yes, I'm pretty sure that was leftover from a last-minute copy-paste on my part. Sorry about that @joezuntz .