Closed GoogleCodeExporter closed 9 years ago
[deleted comment]
I think we only need to show a single +/- column because we are using Bayeselo
in a way that the upper and lower confidence bound should be the same distance
from the mean.
Alternatively, we might want to consider showing the actual upper and lower
confidence bounds rather than their distances from the mean. That is, instead
of <mean> <-> <+>, e.g. "2090 123 123" from the first row of the screenshot, we
could show something like <-> <mean> <+>, e.g. "1967 2090 2213".
Original comment by Jake.McA...@gmail.com
on 5 Oct 2010 at 3:14
I've made the display changes suggest above. I agree that it is more
informative this way. I also ended up changing the value stored in the database
to the actual bound rather than the +/-. This seems generally more useful not
only because that is the value we are always displaying now but also it will
allow us to easily sort based on the lower bound later if desired.
Jake.McArthur has also worked on some javascript whisker plots that should be
quite nice to add after this change goes in.
The new database alter statements are:
ALTER TABLE rankings ADD upper_bound double default '4000' AFTER score;
ALTER TABLE rankings ADD lower_bound double default '0' AFTER score;
Original comment by janzert
on 5 Oct 2010 at 7:46
I forgot to mention in the last comment. The reason bayeselo is currently
giving symmetric bounds is because the accurate confidence interval calculation
had been commented out since we weren't using them anyway before.
The command to calculate better intervals but with a still reasonably fast
calculation is added back into this branch. The best interval calculation
method is cubic with the number of participants (not games played though) so
may be too expensive on the server. It would be nice if this could be tested on
the server to see how long it takes in practice (see note in readgames.py).
Original comment by janzert
on 5 Oct 2010 at 7:51
Original comment by janzert
on 6 Oct 2010 at 4:58
Set review issue status to: WontFix
Original comment by janzert
on 18 Nov 2010 at 7:03
Original issue reported on code.google.com by
janzert
on 5 Oct 2010 at 6:24