|
On June 02 2013 06:43 Grovbolle wrote:Show nested quote +On June 02 2013 06:40 edwahn wrote: Love the new update. I think the offline factor is absolutely the right thing to do. People remember champions for offline tournaments, not online ones!
The Hall of Fame now much more resembles my memory of who I thought were dominant. Thanks for the hard work! The HoF wasn't really taken that much into consideration, we all agreed that it needed som rework. I guess this update had a good side effect on that as well Yeah, I've almost totally forgotten about the HoF, but nice to see that it also got a facelift.
|
How did you alleviate the infinite performances issue?
|
The changes to the system are super sweet. I like the state of the ranking quite a lot.
|
On June 02 2013 06:54 BoZiffer wrote: How did you alleviate the infinite performances issue? Some guy posted in one of my threads once how to do it. It's a couple of months back.
http://static.aligulac.com/method.pdf
Basically just maximise the whole right hand side in equation (1), instead of just maximising the product and then using the update formula.
Edit: Kinda cheating when I said that. Of course performances can still be infinite, but you don't need performances to update ratings.
|
United States8476 Posts
How was the number 1200 derived? Was it mostly an arbitrary number? Or something that you got through trial and error? Or is there a mathematical significance?
|
On June 02 2013 07:45 monk wrote: How was the number 1200 derived? Was it mostly an arbitrary number? Or something that you got through trial and error? Or is there a mathematical significance? Well, I tried tracking the point transfer to the Korean scene. 1200 is definitely on the low end. If you want the most accurate ratings today it should probably be at least 1400. But if I do that, many of the earlier lists start looking really weird. So I landed on 1200 as a compromise.
I conjecture that the best choice is a time-dependent function that starts near 1000 in February 2010 and increases slowly since then, but I haven't really bothered looking into that yet. For example, something like
1000 + L * (1 - exp(-time*rate))
where the question is how to choose the rate, and the limit L.
|
I'm amazed by how much this has been improving, definitely deserves more attention (integration with TLPD maybe?)
I can't say I agree with the "1200 points for Koreans" thing. It seems arbitrary and diminishes Aligulac's value as a totally impartial ranking. I think you've correctly identified the cause of the issue, but I'd like to see something a bit less arbitrary. how about: 1) When a new player enters the system, their starting rating is (some function of) the average rating of their region. Where "region" means country or BNet server. 2) Is it necessary to make any assumption at all about a player's rating before they play any games? Maybe their initial rating could be a function of their first 10 games or something (like placement matches). Maybe there's a way to play with the way ratings deviation changes over time to emulate this effect to some extent. Can points ever be created from nothing, or do people just exchange points with each other? If the latter, maybe make an exception for the first 10 matches to correct inaccuracies with the 1000 assumption.
I think it would help a great deal to have tooltips or little "[?]" links everywhere to explain what everything means. When I see "Leading race: Terran", I'd like to be able to read a summary of what level of play is considered/how it's calculated/etc.
Also, maybe I'm dumb but I can't see a way to go to the page for an event from a player's list of games. This would be useful, I think?
Finally, I'm sceptical that "Most specialised vX" is a sufficiently interesting/important statistic to feature so prominently on the rankings page. Especially since you presumably have to weight it by rating so you dont get random players nobody's heard of.
|
On June 02 2013 14:19 mkwn wrote: When a new player enters the system, their starting rating is (some function of) the average rating of their region. Where "region" means country or BNet server.
I think that could prove unfair for certain countries.
I mean, inflating Koreans makes sense from an empirical standpoint, since their region is by far the most cut-throat and competitive, even down to the ladder.
On the other hand, while your suggestion might seem fairer in theory, it could also underrate players in regions where there's a big discrepancy between average rating and top player(s). I took a look at Norway, just as an example (I'm sure there are better ones still): We have 69 Norwegian players in our database, of which only 10 find themselves at 1000 rating or higher, with an average rating of 860. Do you think it would be fair towards players such as Snute and TargA, simply because their fellow countrymen haven't achieved the same success?
I'm sure BB will try it out at some point, even if just for the sake of it, but IMHO your solution will only create numerous smaller problems for the rating in scenes where there's only a few relatively strong players.
As for BNet servers, we don't have that information because it would be irrelevant since many of the players outside Korea practice on more than one server.
|
On June 02 2013 14:19 mkwn wrote: Especially since you presumably have to weight it by rating so you dont get random players nobody's heard of.
I don't think he does this actually. It has shown some pretty bizarre results at times before the Feb ratings change. The newest ratings adjustments make it even more useful, imo.
Edit: Strange... http://www.aligulac.com/players/45-Soulkey/period/86/ shows that he lost to Jangbi, why does it say he gained 2 points from that? Also, isn't a gain from 1771 to 1774 3 points anyway?
|
On June 02 2013 14:19 mkwn wrote: I think it would help a great deal to have tooltips or little "[?]" links everywhere to explain what everything means. When I see "Leading race: Terran", I'd like to be able to read a summary of what level of play is considered/how it's calculated/etc.
Also, maybe I'm dumb but I can't see a way to go to the page for an event from a player's list of games. This would be useful, I think?
Leading/Lagging: http://aligulac.com/faq/#6
Navigating to events:
Just press the grayed out "premier" area. Any part of the event name will take you to the relevant part of the hierarchy.
|
@MasterOfPuppets:
If all the Norwegian players have a low ranking, and two new unknown Norwegian players named Snute and Targa enter the scene, what do you think is the fairest initial ranking to give them? I think a low ranking is pretty appropriate, without prior knowledge that they'll be good.
Any good player who enters the scene will be initially valued too low. The rating system is designed to quickly push their rating up to match their true skill. The issue at hand is that the pool of points for whole regions is incorrect, and this takes a long time to fix because there is much less interaction between regions than within them.
By BNet server, I meant do it by country but group together countries that have the same server (europe, america, korea, china, sea?) The idea being, try to find some natural classification so that each group is very connected. You could even do some mathsy stuff and decide which countries to group together dynamically, based on the connectedness of the data. Might make an interesting "reports" page too, to see how countries interact.
just a suggestion, anyway.
@Grovbolle:
I mean, on soulkey's page i can see the match history ("Most recent results"), i'd like to be able to click on each match to go to the relevant event. Not a big deal.
|
On June 02 2013 17:13 dcemuser wrote:Show nested quote +On June 02 2013 14:19 mkwn wrote: Especially since you presumably have to weight it by rating so you dont get random players nobody's heard of. I don't think he does this actually. It has shown some pretty bizarre results at times before the Feb ratings change. The newest ratings adjustments make it even more useful, imo. We've done that for a while now.
They're not independent. Think of it like this: part of Soulkey's win against Innovation must have been due to his general skill level, and not just his ZvT skill level. You can get funny-looking results like these if someone plays overwhelmingly against one or two races (like here, 7/8 of Soulkey's games are ZvT).
The numbers are actually 1771.2348961374 (rounded down to 1711) and 1773.7245113775 (rounded up to 1774), so the increase is 2.4896152400999654 (rounded down to 2).
|
On June 02 2013 17:35 mkwn wrote: @MasterOfPuppets:
If all the Norwegian players have a low ranking, and two new unknown Norwegian players named Snute and Targa enter the scene, what do you think is the fairest initial ranking to give them? I think a low ranking is pretty appropriate, without prior knowledge that they'll be good.
Any good player who enters the scene will be initially valued too low. The rating system is designed to quickly push their rating up to match their true skill. The issue at hand is that the pool of points for whole regions is incorrect, and this takes a long time to fix because there is much less interaction between regions than within them.
By BNet server, I meant do it by country but group together countries that have the same server (europe, america, korea, china, sea?) The idea being, try to find some natural classification so that each group is very connected. You could even do some mathsy stuff and decide which countries to group together dynamically, based on the connectedness of the data. Might make an interesting "reports" page too, to see how countries interact.
just a suggestion, anyway.
@Grovbolle:
I mean, on soulkey's page i can see the match history ("Most recent results"), i'd like to be able to click on each match to go to the relevant event. Not a big deal. The "most recent results" currently doesn't link to events. I think it is a matter of not cluttering the page too much, but it should be doable.
|
@mkwn: Those are ideas I'm not even near getting implemented yet, so I'll just say: yeah, I can/will look at them eventually.
|
So, the 1/6 (yesterday) changes made Innovation #1? Sounds reasonable.
edit: This also applies to previous ranking lists right? Sorry if I missed something, but woah
|
On June 02 2013 17:43 neptunusfisk wrote:So, the 1/6 (yesterday) changes made Innovation #1? Sounds reasonable. edit: This also applies to previous ranking lists right? Sorry if I missed something, but woah This also applies to previous lists.
|
On June 02 2013 17:48 Grovbolle wrote:Show nested quote +On June 02 2013 17:43 neptunusfisk wrote:So, the 1/6 (yesterday) changes made Innovation #1? Sounds reasonable. edit: This also applies to previous ranking lists right? Sorry if I missed something, but woah This also applies to previous lists.
Yeah, cool. Innovation #1! :D
|
Ahh, this "point pool" problem is very difficult to work out. Even the very act of assigning players to pools like "foreigner" and "Korean" seems to taint the purity of the system. But, the fact that foreigners play foreigners often and Koreans play Koreans often means (before the recent change) an average Korean would be rated the same as an average foreigner, which is obviously not ideal for rankings. I think theBB is right in that these player pools themselves could be considered to have their own ratings (like right now the Korean pool are rated 1200 and the foreigner pool is rated 1000) but this ratio shouldn't be fixed, rather it should be movable...somehow...
As MasterOfPuppets pointed out, someone like Stephano who is far better than his fellow countrymen might suffer from such a system. I feel that it wouldn't be as bad of an issue as it seems. A foreigner who is truly at the skill level of Koreans would dominate his player pool even harder than in an even system, which helps to make up for his player pool nerf.
Another idea: when a player plays someone from a different pool, increase the significance of the match. I dunno.
E: mkwn hit on much of what I was thinking as well.
|
On June 02 2013 18:14 slowbacontron wrote: Ahh, this "point pool" problem is very difficult to work out. Even the very act of assigning players to pools like "foreigner" and "Korean" seems to taint the purity of the system. But, the fact that foreigners play foreigners often and Koreans play Koreans often means (before the recent change) an average Korean would be rated the same as an average foreigner, which is obviously not ideal for rankings. I think theBB is right in that these player pools themselves could be considered to have their own ratings (like right now the Korean pool are rated 1200 and the foreigner pool is rated 1000) but this ratio shouldn't be fixed, rather it should be movable...somehow...
As MasterOfPuppets pointed out, someone like Stephano who is far better than his fellow countrymen might suffer from such a system. I feel that it wouldn't be as bad of an issue as it seems. A foreigner who is truly at the skill level of Koreans would dominate his player pool even harder than in an even system, which helps to make up for his player pool nerf.
Another idea: when a player plays someone from a different pool, increase the significance of the match. I dunno.
E: mkwn hit on much of what I was thinking as well. The "Koreans start with 1200" solution has helped bridge the "kespa gap" so that the list has caught up now, although it is a bit subjective, the evidence does support this kind of solution.
|
Hmm, are the 1200-rating-Koreans everyone with a Korean nationality in the database? Since then the system would favor players like SeoHyeon and KingKong, too, which is probably not the intention. An additional requirement could be that a player also had to have been in any of the Kespa/ESF teams at some point to slightly narrow down the list of naturally gifted players.
(Hey, I might be a dev too, but I have nothing to do with the rating system, so I can ask stupid questions just like everyone else! )
|
|
|
|