Ticket #7383 (closed: fixed)
Investigating using poisson cost function
Reported by: | Anders Markvardsen | Owned by: | Mike Thomas |
---|---|---|---|
Priority: | major | Milestone: | Release 3.0 |
Component: | Framework | Keywords: | |
Cc: | james.lord@… | Blocked By: | |
Blocking: | Tester: | Anders Markvardsen |
Description (last modified by Anders Markvardsen) (diff)
James has identified some cases where Mantid fit struggles.
Here investigate the potential benefit of using poisson statistics.
See attached python scripts and James writes:
(a) a plain "background" count level in a single group
(b) Asymmetry, with a constant value
(c) a more real example: a slowly relaxing oscillating signal (TF muSR).
Attachments
Change History
comment:1 Changed 7 years ago by Anders Markvardsen
- Description modified (diff)
In addition investigate the effect of the different minimizers. I.e. would changing the minimizer give different results for James' examples
comment:2 Changed 7 years ago by Nick Draper
- Milestone changed from Release 2.6 to Backlog
Moved to backlog at the code freeze for R2.6
comment:4 Changed 7 years ago by Anders Markvardsen
- Summary changed from Investigate examples where Mantid fit fails to Investigating using poisson cost function
comment:6 Changed 7 years ago by Mike Thomas
The python script (a) was adapted into a matlab script and a fit using poisson cost function added. Poisson data is generated according to a count rate and each cost function is used to calculate the background rate. The difference between the two calculated rates and the rate used to generate the poisson data is plotted, showing clear regions where the poisson cost function is better than the gaussian based least squares cost function. The scripts and functions are attached(fitLowStats.m, flatBG.m, flatBGpo.m), as are the result plots (flatBGfigs.png). The results suggest it is necessary to implement the cost function in Mantid Fit. See ticket #7914
comment:7 Changed 7 years ago by Mike Thomas
- Owner changed from Anders Markvardsen to Mike Thomas
- Status changed from new to verify
- Resolution set to fixed
comment:8 Changed 7 years ago by Anders Markvardsen
- Status changed from verify to verifying
- Tester set to Anders Markvardsen
comment:9 Changed 7 years ago by Anders Markvardsen
- Status changed from verifying to closed
As mike showed me attached data show poisson give at least as good as least squares for higher counts and better for lower count data
comment:10 Changed 5 years ago by Stuart Campbell
This ticket has been transferred to github issue 8229