| Forum Index > SC2 General |
|
darmousseh
United States3437 Posts
On October 25 2010 17:20 Slaytesics wrote: also, this equation might be helpful in determing the actual ranking system and equation At at around 2:50 I hadn't seen this. This is definitely almost identical to the trueskill formula. Awesome. | ||
Excalibur_Z
United States12235 Posts
On November 17 2010 05:46 darmousseh wrote: I hadn't seen this. This is definitely almost identical to the trueskill formula. Awesome. That's not for matchmaking though, that's the calculation to determine the racial adjusted win percentages. | ||
|
darmousseh
United States3437 Posts
On November 17 2010 05:51 Excalibur_Z wrote: That's not for matchmaking though, that's the calculation to determine the racial adjusted win percentages. Looking more at this formula it is quite retarded. The 3 variables in the probability are nowhere to be found in the formula itself, so the variables on the right side are probably derived. Instead of this being a probability function it is more likely to be a distribution function acounting for race and opponents race in the equation, however, because the system does not differentiate your ratings based on matchup, they are including all of the games played. The 3 variables in the distribution(gamma, theta and psi) are usually used to represent 3 different variables to determine the distribution. One of them is most certainly skill, the other 2 are probably race and matchup (in trueskill they use skill, team, and team performance). The denonimator is very similar to the message passing equation from trueskill, and the numerator looks like a standard bayesian probabilty seperating the influences out. So what is probably happening is that they are determining the skill distribution of each matchup while factoring out the matchup influence (easily done using bayesian probabilty) and then comparing these distributions mean and sigma to determine the win % (which is why the equation is nice and huge). Other than that, i hate that blizzard doesn't just say what it is. The community is not retards, we went to college, we got degrees in math and science. The final thing is that i do not believe determining a distribution is possible. If even one of those variables listed is dependant on another variable, then using trueskill is not an accurate way to determine skill (IE if performance is highly dependant on matchup). In chess, they can get away with trueskill because the number of each matchup you play is almost 50/50 (50% of the time you play white vs black, 50% you play black vs white). From that they can detemine a players skill overall. However, because of the influence of matchup to skill in starcraft determining real skill is not possible. | ||
|
vanick
United States53 Posts
If you meant player skill (which I think you probably did not) then you're saying bayesian inference systems are not good at approximating skill which is pretty demonstrably not the case. If you meant the racial win % then what it seems like you're saying is that since most players only play a single race that skews things because their performance is not constant against various opponents. That is, Player A usually plays Terran, and is better in TvP than in TvZ. So what you're saying then is that because different players each have this difference (and some more than others) that completely confounds the calculations Blizzard does. This is incorrect for a couple reasons. First of all, if this pattern existed then Blizzard would have the data to model it and control for it. Second of all, if this pattern existed strongly in aggregate (if most terrans were better at TvP than at TvZ, say) then that would also show up in the win percentages, because the TvP win % would be higher than it ought to be. | ||
|
ipwnN00bz
United States42 Posts
Im sure Idra played a ton of games, and likely the system got very "sure" (ie the sigma value approached 0) that he belonged in whatever league he was trying to get out of - maybe because of the ranks of the people he was up against, etc. This is a common problem in learning machines that use a momentum factor. Without time taken into account i dont think the system you are describing can be guaranteed to converge. The easiest solution to this would be to let uncertainty build up over time, then become more accurate, and repeat this in cycles (ie simmulated annealing).Losing games would raise the uncertainty (probably more rapidly than letting it build up over time) and allow you to move out of the local minimum. It would negatively affect your MMR but you could "move around" more easily as the system becomes more uncertain. Basically if you play 18 hours a day and run your sigma up, loosing is probably faster than waiting for it to go back down over time. Just a guess. | ||
|
Excalibur_Z
United States12235 Posts
On November 20 2010 09:26 ipwnN00bz wrote: If i had to guess as to the reason behind the "needing to loose to rank up" phenomenon i would say that it doesnt have anything to do with a "moving average" but that the system is getting stuck at a local minimum. Im sure Idra played a ton of games, and likely the system got very "sure" (ie the sigma value approached 0) that he belonged in whatever league he was trying to get out of - maybe because of the ranks of the people he was up against, etc. This is a common problem in learning machines that use a momentum factor. Without time taken into account i dont think the system you are describing can be guaranteed to converge. The easiest solution to this would be to let uncertainty build up over time, then become more accurate, and repeat this in cycles (ie simmulated annealing).Losing games would raise the uncertainty (probably more rapidly than letting it build up over time) and allow you to move out of the local minimum. It would negatively affect your MMR but you could "move around" more easily as the system becomes more uncertain. Basically if you play 18 hours a day and run your sigma up, loosing is probably faster than waiting for it to go back down over time. Just a guess. It's definitely a moving average, we had that confirmed at Blizzcon. Your post basically outlines something similar to our initial theory, which was that it's much closer to TrueSkill. That's actually not the case, and sigma actually never shrinks that much in SC2. We were proven wrong at Blizzcon and adjusted the OP accordingly. | ||
|
sqrt
1210 Posts
| ||
|
Nimic
Norway1360 Posts
On November 21 2010 20:18 sqrt wrote: Wait so...if I am Bronze in Team and Platnum in 1v1, does my team rating affect my solo performance? I have no real knowledge of this, but it seemed like it to me particularly at one point. I'm Platinum 1v1, not really having played much, very low on points. At the same time I'm Diamond 2v2, decently high rated, as 2v2 goes. At one point my 2v2 did very well over a couple of nights gaming, and the next time I played 1v1 I seemed to be put up against higher ranked people than I would otherwise be. I was maybe 500 Platinum at the time, and was put up against 1100-1500 diamonds (and some Plats). I met some pretty good people and had a pretty bad streak, and suddenly it felt like we were facing easier opponents in 2v2. Of course, I can very much see the possibility that this is just something I've made up in my head, and made the events fit better and have more of a connection (if any) than they really do. Damn you, human brain... So you could take my anecdotal and admittedly uncertain story, or get a straight answer from someone who actually knows. Let's hope for the second one. | ||
|
sqrt
1210 Posts
On November 21 2010 20:31 Nimic wrote: I have no real knowledge of this, but it seemed like it to me particularly at one point. I'm Platinum 1v1, not really having played much, very low on points. At the same time I'm Diamond 2v2, decently high rated, as 2v2 goes. At one point my 2v2 did very well over a couple of nights gaming, and the next time I played 1v1 I seemed to be put up against higher ranked people than I would otherwise be. I was maybe 500 Platinum at the time, and was put up against 1100-1500 diamonds (and some Plats). I met some pretty good people and had a pretty bad streak, and suddenly it felt like we were facing easier opponents in 2v2. Of course, I can very much see the possibility that this is just something I've made up in my head, and made the events fit better and have more of a connection (if any) than they really do. Damn you, human brain... So you could take my anecdotal and admittedly uncertain story, or get a straight answer from someone who actually knows. Let's hope for the second one. Well, my 2v2 partner may be looking for another person then. | ||
|
Excalibur_Z
United States12235 Posts
On November 21 2010 20:18 sqrt wrote: Wait so...if I am Bronze in Team and Platnum in 1v1, does my team rating affect my solo performance? Only for the first placement match in a new team or bracket. For example, Vanick and I were in a Diamond 2v2 team but I hadn't played any 1v1. When I did start 1v1, I was put against Diamond-strength opponents from the very beginning, and that's because it was carrying over my 2v2 performance into 1v1. After the first placement match, the MMRs are completely separate. | ||
|
neo_rtr
Sweden70 Posts
Some stats, iam 1 or 2 on my division 391 won, 357 lost, ratio of 52.27% | ||
|
ipwnN00bz
United States42 Posts
| ||
|
ipwnN00bz
United States42 Posts
| ||
|
Excalibur_Z
United States12235 Posts
On November 23 2010 01:00 ipwnN00bz wrote: I watched the video of the Blizzcon discussion (including your question)- the bit about the "moving average" was ambiguous at best- it could have just been simplifying what is essentiall a graduate level computer science topic to a more approachable concept. Considering the subject matter and the depth that we were discussing with him after the panel ended, I don't think this is the case. | ||
|
iamtheone
15 Posts
I clearly no longer belong in Bronze. I'm still stuck here  I think it might have something to do with being crap in 1v1 at the beginning. I honed my skills in 4v4, went back to 1v1 and started slaughtering most opponents. | ||
|
ThaZenith
Canada3116 Posts
He was in his third placement match in 1v1, and was "favored" even though he has lost his first 2 matches. He was gold in 2v2 with 200+ games played. A 0-2 player favored over a maybe 30-13 (at the time) player seems silly otherwise. | ||
|
Excalibur_Z
United States12235 Posts
On November 27 2010 11:28 ThaZenith wrote: I can almost guarantee that your rating in team matches like 2v2, 3v3, 4v4, have an impact on your 1v1 matchmaking. He was in his third placement match in 1v1, and was "favored" even though he has lost his first 2 matches. He was gold in 2v2 with 200+ games played. A 0-2 player favored over a maybe 30-13 (at the time) player seems silly otherwise. We knew that already. When you start a new team or bracket it uses data from your other brackets as a starting point. | ||
|
iamtheone
15 Posts
It would be better if I was to consistently play against better players and consistently lose and improve my skill, rather than play a better player, lose the win the next couple of matches because I'm better than the next batch it sets me up against. Thoughts? | ||
|
theSAiNT
United States726 Posts
| ||
|
DMac4400
Canada1 Post
I bought the game and really had no idea what was going on. I lost all my preliminary matchups and than lost nine straight games. I got some help from friends and really picked up my game play. I am now 37 - 25 in a Bronze league playing and beating Silver League players.I have now won 15 in a row. Am I looking at a promotion or am I winning too much for the sigma to catch up? | ||
| ||
Road to EWC
DreamHack Dallas Group Stage
[ Submit Event ] |
|
|
|
Afreeca Starleague
BeSt vs Soulkey
AllThingsProtoss
Road to EWC
BSL: ProLeague
Cross vs TT1
spx vs Hawk
JDConan vs TBD
Wardi Open
SOOP
NightMare vs Wayne
Replay Cast
Replay Cast
GSL Code S
Cure vs Zoun
Solar vs Creator
The PondCast
[ Show More ] |
|
|
EPT
