EPTZerg rule finally toppled - Page 5
| Forum Index > SC2 General |
|
delHospital
Poland261 Posts
| ||
|
Grovbolle
Denmark3804 Posts
On April 19 2013 01:12 delHospital wrote: I wonder if you could extend the system with some form of backward propagation of skill, something like "TrueSkill through time". I think this could help with the KeSPA problem, where a bunch of players were added to the pool that are on a level much higher than the initially assumed 1000 points. I do believe that is already used. (I think it's called forward filtering backwards smoothing or something) | ||
|
Goolpsy
Denmark301 Posts
So a player winning many games against 1300 Europeans, might earn 'quite alot of point' maybe getting 1500? But if the same player won a few GSL games (vs players with 1700 points) the point climb would be steeper. => Winning GSL games gives you more points, because you fight better opponents. (Note: Beating Life at MLG or GSL, or even an online tournament, earns you the same amount of points; which is kind of fair?) Question though: You get a different amount of point from beating an opponent 4-0, instead of 4-3, right? | ||
|
neozxa
Indonesia545 Posts
| ||
|
Grovbolle
Denmark3804 Posts
On April 19 2013 01:15 Goolpsy wrote: Again, as stated several times already; The amount of points you get, depends on the 'level'/'score'/'rating' of the players you are playing against. So a player winning many games against 1300 Europeans, might earn 'quite alot of point' maybe getting 1500? But if the same player won a few GSL games (vs players with 1700 points) the point climb would be steeper. => Winning GSL games gives you more points, because you fight better opponents. (Note: Beating Life at MLG or GSL, or even an online tournament, earns you the same amount of points; which is kind of fair?) Question though: You get a different amount of point from beating an opponent 4-0, instead of 4-3, right? Yes. Actually you can gain points by losing as well. Say that Life beats "insert random nobody" 4-3 Life was expected to beat him 4-1 (Which is called expected score here http://aligulac.com/players/3-Life/period/82/). The random nobody then gets a few points which life loses, because life underperformed compared to what was expected of him in a Bo7 | ||
Zealously
East Gorteau22261 Posts
Thanks for the hard work, TheBB and crew. As long as Life is #1, I won't complain! | ||
|
TheBB
Switzerland5133 Posts
Should tournaments be weighted? A good mathematical model has few parameters. The standard Glicko system makes do with four (period length, initial uncertainty, uncertainty floor and uncertainty decay rate). Having experimented for a while I have decided to add a fifth, the factor by which offline matches should be weighted relative to online ones. (This change is not live yet, by the way. I'm still working on it.) If I were to start weighing tournaments it would introduce something like a hundred additional parameters. At this stage it's no longer a useful model. Some condensation is necessary. For example: you can do linear weighing by prize pool, or you can do linear weighing by the mean rating of participating players, or something like that. I'm not saying this will be useless, but conventional mathematical wisdom is to enhance your model with the parameters that are likely to yield the best results with the least expense in complexity first, and while I can't out of hand dismiss the benefits of this scheme, I can tell that it will be complex. Also, please note that you already get more points by beating higher rated players. As such, wins in the GSL automatically count for more because you will likely face better opposition there. Also worth keeping in mind is that if I increase the weight of a tournament, it will not only increase the point gain of the winners, but also the losses of the losers. It will not make the mean rating of Code S higher, but it would increase the gap between the best and the worst Code S players. This may not be what you actually want. Filtering I have already implemented backwards filtering, but we're still discussing where it makes sense to display filtered numbers, and where it makes sense to display the raw ones. Currently, only the player charts show filtered data. I was reluctant to use those in the rating lists, because then they can change after the fact, which is somewhat counterintuitive and can confuse users (and I'm not short of confused users, as you can see). If people want to see filtered data in rating lists, then I can do that. Note that it will not change the current list though, only the past ones. | ||
|
Kasaraki
Denmark7115 Posts
| ||
|
Grovbolle
Denmark3804 Posts
On April 19 2013 01:30 Kasaraki wrote: Everytime Aligulac content is posted there's a handful of really stupid comments either not reading OP or the faq, or lacking basic grasp of math. >_< Please don't let that discourage you guys, personally I find Aligulac really useful and very interesting. I always feel like the few of us who actively post in these threads, explaining the site etc. ends up coming off as ULTRA defensive and not willing to take criticism. (Me, Conti, TheBB and Otolia mainly). Constructive criticism is always good, however before you start questioning the math, please remember this. TheBB is studying a PhD in math. So please bring at least a decent argument and get familiarized with Glicko rating and basic statistical knowledge as well as what the site actually is and most importantly, what it is NOT before derping out stuff like: "No KESPA = Shit rating" | ||
|
KillerDucky
United States498 Posts
| ||
|
TheBB
Switzerland5133 Posts
On April 19 2013 01:34 Grovbolle wrote: I always feel like the few of us who actively post in these threads, explaining the site etc. ends up coming off as ULTRA defensive and not willing to take criticism. (Me, Conti, TheBB and Otolia mainly). Yes, please, that is not my intention at all. I love the fact that this sparks some debate. | ||
|
Novacute
Australia313 Posts
| ||
|
Godwrath
Spain10107 Posts
On April 19 2013 01:55 Novacute wrote: A top 10 foreigner list that does not include BabyKnight and Mana? There's definitely something wrong here. Maybe, take your time to read the OP, the FAQ and several explanations in this thread ? FUCK IT. IT SUCKS. Btw, there was a time i used to "cheat" using aligulac on liquibets. It got me to top 20 easilly haha. | ||
|
Jerubaal
United States7684 Posts
| ||
|
delHospital
Poland261 Posts
Do you think it would change much if you seeded Koreans with higher initial ratings that foreigners? I realize that it doesn't fit nicely with the model, but would it be more accurate? | ||
|
Novacute
Australia313 Posts
On April 19 2013 01:57 Godwrath wrote: Maybe, take your time to read the OP, the FAQ and several explanations in this thread ? FUCK IT. IT SUCKS. Btw, there was a time i used to "cheat" using aligulac on liquibets. It got me to top 20 easilly haha. If a player has not played any games for four periods (eight weeks, or about two months), they will not be included. Mana has been playing very well in the ATC, and he's been a prominent player in recent EU tournmaents like Zotac and WCS europe. Surely, that is less than eight weeks. Edit: Found him, surprised he wasn't ranked any higher. I clearly overestimated his ability, but the way it's gauged is very arbitrary, he's been consistently winning against Koreans, shouldn't that yield a far better result than otherwise indicates? | ||
|
TheBB
Switzerland5133 Posts
On April 19 2013 02:00 Novacute wrote: If a player has not played any games for four periods (eight weeks, or about two months), they will not be included. Mana has been playing very well in the ATC, and he's been a prominent player in recent EU tournmaents like Zotac and WCS europe. Surely, that is less than eight weeks. http://aligulac.com/periods/83/?page=2&sort=&race=ptzrs&nats=all Mana is on 72nd place in the world, and 17th place among foreigners. He has not been removed, he is simply not in the top 10 on account of there being limited space there (namely, 10 spots). I hope you didn't think that only 20 people have played a game in the last two weeks. I'm not about to copy the whole list of 680 people in here. | ||
|
ratbert
Germany1041 Posts
| ||
|
KrazyTrumpet
United States2520 Posts
| ||
|
Aeceus
United Kingdom1278 Posts
Your a legend  | ||
| ||
