[REDMINE1D-36] [RM-6165] use likelihood instead of posterior for merit Created: 04/Jun/21 Updated: 13/Jun/23 Resolved: 13/Jun/23 |
|
| Status: | Done |
| Project: | 1D Redmine |
| Component/s: | None |
| Affects Version/s: | None |
| Fix Version/s: | None |
| Type: | Task | Priority: | Normal |
| Reporter: | Redmine-Jira Migtation | Assignee: | Redmine-Jira Migtation |
| Resolution: | Done | Votes: | 0 |
| Labels: | None | ||
| Remaining Estimate: | Not Specified | ||
| Time Spent: | Not Specified | ||
| Original Estimate: | Not Specified | ||
| Attachments: |
|
| Description |
|
Created on 2020-12-04 08:51:23 by Didier Vibert. % Done: 100 we should try to integrate the likelihood under the peak instead of the posterior, ie not renormalize to unit integral, that is not divide by the evidence which encode the probability of the data (goodness of fit). This has not impact on the ordering of candidates for one spectra, but may help as a confidence indicator from one spectra to the other. In the purity/completness we could as well try to use the evidence (ie the likelihood integrated over the whole z range) |
| Comments |
| Comment by yuki.moritani [ 13/Jun/23 ] |
|
Comment by Didier Vibert on 2021-11-29 10:33:59: sur la simulation ELcosmos, en remplaçant le merit (prob intégrée) par la vraisemblance intégrée (en log) on ne discrimine plus du tout entre les spurious et les targets: note: le calcul du nouveau merit est le suivant: les valeurs des évidences sont très étalées (range de taille exp(300)...):
Du coup, le merit (ie la fraction du pic sur le total de la pdf ) est noyé complètement par la valeur de l'évidence qui elle n'est pas du tout discriminante. |
| Comment by yuki.moritani [ 13/Jun/23 ] |
|
Comment by Didier Vibert on 2021-11-29 10:35:44: |