Re: [Starpu-devel] Problem with performance model calibration

[Samuel Thibault <samuel.thibault@inria.fr>]

Mirko Myllykoski, le mar. 09 juin 2020 13:54:25 +0200, a ecrit:
> That is going to cause problems due to the log(y[i]-c) statement in the
> test_r, compute_b and compute_a functions (negative argument for log).

Oh, right!

Actually, these values shouldn't be taking into account in the linear
regression being the non-linear regression: we are regressing y[i]-c,
not y[i], and that difference will *not* get smaller.

I have fixed this and other issues in the different branches:

- ignore the values close to c
- start from the first fourth of the values, instead of the minimum
- weight the values by their number of measurements, otherwise you'd get
  disturbed by small outliers.

Could you try that with your application?  I believe the result should
be much better.

Samuel