Arimaa Forum - Print Page


    
      
        Arimaa Forum
        (http://arimaa.com/arimaa/forum/cgi/YaBB.cgi)
      

        Arimaa >> Site Discussion >> Link to "recent upsets"
        
(Message started by: obiwan on Jul 22^nd, 2006, 10:09pm)

Title: Link to "recent upsets"
Post by obiwan on Jul 22^nd, 2006, 10:09pm

In a game comment, I suggested a "recent upsets" link which would show games where the winner was rated lower than the loser. Ron Weasley and Fritzlein liked the idea so I thought I'd suggest it in this forum.

Title: Re: Link to "recent upsets"
Post by Fritzlein on Jul 23^rd, 2006, 2:04pm

Yes, I really like the idea. (Also it is well to post it here where Omar will see it eventually.) Maybe there should be a minimum of 100 points rating difference so doesn't display a bunch of games between basically equal people.

Title: Re: Link to "recent upsets"
Post by DorianGaray on Jul 23^rd, 2006, 4:52pm

I'd say 200 points at least.

Title: Re: Link to "recent upsets"
Post by arimaa_master on Jul 24^th, 2006, 11:27am

Ok, lets say 150 points seem to be reasonable compromise :)

Title: Re: Link to "recent upsets"
Post by seanick on Jul 28^th, 2006, 11:09am

Out of May 2006 human vs human games, (admittedly a very limited sample but I am late for work:)):

where the loser is rated higher than the winner, these are the percentages:

100+ 43.6%
200+ 39.3%
300+ 24.8%
400+ 12.8%
<my vote for an upset is somewhere between these two>
500+ 5.1%
600+ 0.9%

Title: Re: Link to "recent upsets"
Post by Ryan_Cable on Jul 28^th, 2006, 11:48am

Wow, it would be interesting to see if those numbers held up over a longer time scale. The ELO equation predicts substantially different results:

50 0.4285
100 0.3599
150 0.2966
200 0.2402
250 0.1916
300 0.1509
350 0.1176
400 0.0909
450 0.0697
500 0.0532
550 0.0404
600 0.0306
650 0.0231
700 0.0174
750 0.0131
800 0.0099

Title: Re: Link to "recent upsets"
Post by chessandgo on Jul 28^th, 2006, 1:29pm

well, this doesn't seem substantially different :) prediction is just a bit too low for little upsets and too high for big upsets, but figures are comparable, aren't they ?

Title: Re: Link to "recent upsets"
Post by Fritzlein on Jul 28^th, 2006, 1:36pm

on 07/28/06 at 11:48:54, Ryan_Cable wrote:

The ELO equation predicts substantially different results:

No, I don't think that is what seanick is measuring. He is reporting out of games that actually were upsets. Those results are clearly contingent on what games have been played. If, for example, no games are ever played between opponents within 100 rating points of each other, then 100% of upsets will have a gap of at least 100 points, etc.

He is specifically not saying that a gap of 400 points results in 12.8% upsets, but rather that 12.8% of the upsets that actually occurred had a gap of 400 points or more.

Title: Re: Link to "recent upsets"
Post by Ryan_Cable on Jul 28^th, 2006, 2:14pm

100+ 44.7
200+ 75.5
300+ 192.7
400+ 333.3
500+ 507.9
600+ 816.7

Those are the rating differences corresponding to seanick’s upset percentages. I don’t expect the ratings to be particularly accurate for 600+, but I am quite surprised by the other results. To my mind, a rating error of more than 100 points for aggregate data like this is quite substantial.

Also, I am surprised by the direction of the error. I have argued elsewhere in the forum that the ratings are too compressed and would spread out more over time, but this seems to indicate that the ratings are actually too spread out.

I would like to see the results for a larger period of time and also for data restricted to 1600+ or 1700+ HvH games, given that ratings near 1500 tend to be particularly inaccurate.

Title: Re: Link to "recent upsets"
Post by Ryan_Cable on Jul 28^th, 2006, 2:22pm

on 07/28/06 at 13:36:28, Fritzlein wrote:

He is specifically not saying that a gap of 400 points results in 12.8% upsets, but rather that 12.8% of the upsets that actualy occurred had a gap of 400 points or more.

Oh, then nothing I just wrote makes any sense. :-) Still, I would be interested to see an analysis that did it that way.

Title: Re: Link to "recent upsets"
Post by seanick on Jul 28^th, 2006, 10:44pm

ok, it turns out I ran into a bug in excel. some of the sort's I did resulted in only some of the columns getting sorted correctly. so some of the games I had based my estimates on were not hvh, and so those previous numbers are complete bunk.
this time I was more careful to make sure that didn't happen, and I came up with much more reasonable numbers.

first, only 19% of the games, TOTAL, end with the loser having a higher rating ..
they break down like this, out of only 116 games of hvh during May.

sorry for the confusion...

0 23 19.83%
100 16 13.79%
200 9 7.76%
300 3 2.59%
500 2 1.72%

Title: Re: Link to "recent upsets"
Post by seanick on Jul 28^th, 2006, 10:59pm

since those games are actually the upsets for the month of may I figured I would just provide that data as well....

here is the set of games for winners rated 100+ points less than the loser:

id	wusername	wrating	busername	brating	diff.	result
31546	blue22	2019	Belbo	1908	111	b
32363	Ryan_Cable	2130	chessandgo	2015	115	b
31998	chessandgo	1958	PMertens	2083	125	w
32221	chessandgo	1999	Ryan_Cable	2134	135	w
32244	KT2006	1678	seanick	1535	143	b
32350	Calumet45	1402	seanick	1558	156	w
32380	Karlo	1657	nbarriga	1500	157	b
32433	blue22	1997	jdb	1796	201	b
31938	robinson	2177	Belbo	1958	219	b
31556	Fritzlein	2345	Adanac	2120	225	b
32175	Fritzlein	2319	PMertens	2061	258	b
32441	chessandgo	2018	kamikazeking	1745	273	b
30834	chessandgo	1797	unic	1521	276	b
31485	chessandgo	1794	Adanac	2141	347	w
31588	chessandgo	1808	Fritzlein	2329	521	w

Title: Re: Link to "recent upsets"
Post by chessandgo on Jul 29^th, 2006, 12:12am

heh, I'm very proud to have the 2 biggest upsets in both directions 8)
I fear I'll still have the record for the worst performance this month :P

Title: Re: Link to "recent upsets"
Post by Fritzlein on Jul 29^th, 2006, 12:58am

That list of games is an interesting reminder that chessandgo's rating rose 200 points in May. :o Based on this list of games it seems that we would have plenty to talk about if we listed only upsets with a rating difference of 200 points or more, including HvB and BvB games.

Title: Re: Link to "recent upsets"
Post by DorianGaray on Jul 29^th, 2006, 5:47am

on 07/29/06 at 00:58:40, Fritzlein wrote:

we could put it in lieu of "human games" which is now redundant with "recent games" (given the new features).

I think that the only interesting list is "upset HvH games".

(Bots can be won using recipes, humans can't)

Title: Re: Link to "recent upsets"
Post by Fritzlein on Jul 29^th, 2006, 8:58am

on 07/29/06 at 05:47:57, DorianGaray wrote:

we could put it in lieu of "human games" which is now redundant with "recent games" (given the new features).

Good point: we could remove the redundant link to avoid clutter.

Quote:

I think that the only interesting list is "upset HvH games". (Bots can be won using recipes, humans can't)

Perhaps it wouldn't be too hard for Omar to add the same game type selection feature to a "recent upsets" page as already exists for the "recent games" page. If new folks start developing bots, I'll be quite interested in BvB upsets too.

Title: Re: Link to "recent upsets"
Post by obiwan on Jul 31^st, 2006, 10:57pm

on 07/29/06 at 05:47:57, DorianGaray wrote:

I'm rather proud of my recipe vs clueless. I think HvB upsets might be interesting. It would suggest a new player becoming stronger or developing a new recipe.