[ 
https://issues.apache.org/jira/browse/LUCENE-6894?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
 ]

Paul Elschot updated LUCENE-6894:
---------------------------------
    Attachment: LUCENE-6894.patch

Patch of 15 Nov 2015.
Resolve conflicts after LUCENE-6276.

I tried this patch with the wikimedium5m benchmark. This showed no significant 
differences to current trunk, the differences where never bigger than 2.1% 
either way, and well within the standard deviations.
This could be because the patch here should have influence for more complex 
queries than the one in the benchmark. I might try to add more complex queries 
to the benchmark later.

> Improve DISI.cost() by assuming independence for match probabilities
> --------------------------------------------------------------------
>
>                 Key: LUCENE-6894
>                 URL: https://issues.apache.org/jira/browse/LUCENE-6894
>             Project: Lucene - Core
>          Issue Type: Improvement
>          Components: core/search
>            Reporter: Paul Elschot
>            Priority: Minor
>         Attachments: LUCENE-6894.patch, LUCENE-6894.patch
>
>
> The DocIdSetIterator.cost() method returns an estimation of the number of 
> matching docs. Currently conjunctions use the minimum cost, and disjunctions 
> use the sum of the costs, and both are too high.
> The probability of a match is estimated by dividing available cost() by the 
> number of docs in a segment.
> The conjunction probability is then the product of the inputs, and the 
> disjunction probability follows from De Morgan's rule:
> "not (A and B)" is the same as "(not A) or (not B)"
> with the probability for "not" computed as 1 minus the input probability.
> The independence that is assumed is normally not there. However, the cost() 
> results are only used to order the input DISIs/Scorers for optimization, and 
> for that I expect this assumption to work nicely.



--
This message was sent by Atlassian JIRA
(v6.3.4#6332)

---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]

Reply via email to