EPT[Patch 4.14] Gnar General Discussion - Page 53
| Forum Index > LoL General |
|
wussleeQ
United States3130 Posts
| ||
|
Anakko
France1934 Posts
| ||
|
Crusnik
United States5378 Posts
On August 22 2014 08:29 Anakko wrote: Nice art, but the heels for Vayne are kinda ridiculous... Like who the fuck fights with heels lol Especially considering how she fights, I can't believe she actually hunts demons with heels and doesn't end up a snack. | ||
|
Goumindong
United States3529 Posts
On August 22 2014 07:50 Mylax wrote: Why would you even want to analyze stuff like KDA and DF? It has no influence on your decision making whatsoever. And why wouldn't it work to find the best linear combination to predict the outcome of the games? What is the endogeneity problem you are talking about? It's obvious that the KDA is correlated with the outcome of the game. Are you saying that the KDA is correlated with the role you play and therefore under/over estimates the "real" significance of KDA in relation to outcome of the games? That only makes sense if the role you play is correlated with the outcome of the games. But we are talking about aggregated data, so I don't quiet understand your point. In stats the correlation has to go "one way" or the math is "inconsistent". By inconsistent i mean that if you increase the amount of data your estimator does not go towards the actual value. Generally this is called an endogeneity problem, because the wrong variable is endogenous. I.E. y=f(x) and x=g(y). In layman's terms you might say there is a feedback loop. If your model is y=xB+e then x=g(y) implies that x correlates with the error term. x'e =/= 0 A simply way to think of this is to write out the solution to the linear estimator. Lets solve: y=xB+e -> x'y=x'xB+x'e x'e= 0 because x is exogenous (not endogenous to y) so we finish easily. (x'x)^(-1)x'y=B_est But if x'e isn't zero then B_est = (x'x)^(-1) x'(y-e) Which we cannot do because we don't have e, our math relies on x'e=0. Indeed we set the error values such that x'e=0 So you can't just look at a linear combination of KDA in order to predict wins. It doesn't work like that. You have to find an instrumental variable for K,D, and A such that this variable(s) correlate with K,D, and A and wins but does not correlate with the error term (that is there is no x=g(y) for our instrumental variable) I can be more precise with the math (at least in the simple linear 2 stage case) but i won't be doing it for the generalized case. The general trick is that by putting a second linear regression in between the two we get another variable which zeroes out the error term, returning us to consistency. For a more technical explanation you can see the wiki. edit: Basically the problem is that winning the game leads to more kills and less deaths just by being ahead. So a simple structure obscures the effect that kills/deaths/assists have on winning which means we will incorrectly estimate a best linear combination to use as a predictor. | ||
|
Kenpark
Germany2350 Posts
So is Shiphtur just rly rly overrated ? Most of the season mid was basicaly 1on1 for the first 10 mins and Shiphtur is behind or even (+5-5) in over 90 % of games. Ofc champion pools play a role, but man didnt expect this at all. | ||
|
Gorsameth
Netherlands22131 Posts
On August 22 2014 09:14 Kenpark wrote: http://www.ongamers.com/articles/cs-10-minutes-a-deep-analysis-of-na-lcs-lane-phase/1100-2123/ So is Shiphtur just rly rly overrated ? Most of the season mid was basicaly 1on1 for the first 10 mins and Shiphtur is behind or even (+5-5) in over 90 % of games. Ofc champion pools play a role, but man didnt expect this at all. Thought that became apparent when his lane opponents have the highest cs at time x over and over again. even when playing something strong like Ziggs the guy is just terrible at the cs game. | ||
|
cLutZ
United States19574 Posts
On August 22 2014 09:14 Kenpark wrote: http://www.ongamers.com/articles/cs-10-minutes-a-deep-analysis-of-na-lcs-lane-phase/1100-2123/ So is Shiphtur just rly rly overrated ? Most of the season mid was basicaly 1on1 for the first 10 mins and Shiphtur is behind or even (+5-5) in over 90 % of games. Ofc champion pools play a role, but man didnt expect this at all. Personally I think Crumbz is terrible and Shiptur is a player who already errs on the side of caution. So if you have a bad jungler who always needs to babysit your botlane, and are naturally cautious...the result is pretty easy to predict. I do think his over-caution is a weakness that teams can exploit though. Curse, C9, and CLG come to mind because those mids don't really try (or if you are Voy, are incapable of) to dominate 1v1. So someone like Hai is like "sweet, no pressure, looks like your botlane is dead!" Or voy is like "sweet, no pressure, my terribleness is hidden!" | ||
|
IMoperator
4476 Posts
| ||
|
Mylax
Germany21 Posts
Since winning is positively correlated with getting more kills, a simple linear regression will underestimate the effect of kills on winning. But I never thought of the linear combination as something like that. It isn't supposed to tell you how big the effect of a kill on the game is. All it tells you is that a specific combination of the kills, deaths, assists over a million games predicted the outcome of the game the best. I.e. high KDA often led to win and low KDA to loss. The specific ratio doesn't tell you anything about each component. I will try to give you an example (Numbers made up): if I count every kill as 1.38 points, every death as 1.23 points and every assist as 3.23 points the resulting number will be the best at predicting the outcome of the game. The specific numbers in my example for each component can't be interpreted. | ||
FakeSteve[TPR]
Valhalla18444 Posts
...They buffed MF so much that even shitcombo can win with her? rito? | ||
|
Goumindong
United States3529 Posts
On August 22 2014 09:39 Mylax wrote: I understand what you are saying in the context of the linear regression. Basically you want to find numbers that say: "every kill increases the likelihood to win the game by x%", correct? Since winning is positively correlated with getting more kills, a simple linear regression will underestimate the effect of kills on winning. But I never thought of the linear combination as something like that. It isn't supposed to tell you how big the effect of a kill on the game is. All it tells you is that a specific combination of the kills, deaths, assists over a million games predicted the outcome of the game the best. I.e. high KDA often led to win and low KDA to loss. The specific ratio doesn't tell you anything about each component. I will try to give you an example (Numbers made up): if I count every kill as 1.38 points, every death as 1.23 points and every assist as 3.23 points the resulting number will be the best at predicting the outcome of the game. The specific numbers in my example for each component can't be interpreted. Its (almost) the exact same math. In the case of this linear combination we would probably use a logistic regression in order to avoid nonsensical results. But they have the same endogeneity problems as the strictly linear case. Its just easier to talk about the strictly linear case when explaining the basic concept. edit 3: its also significantly easier to deal with the problem in the simple linear case edit1: take the linear case again. I wrote y=xB+e this has has y as a 1xn vector. x as a nxk matrix and B has a kx1 vector. Or y= kA + dB + aC + error that last part is our linear combination and that combination will be the best linear combination to predict Y. The problem is that because Y is 0 or 1 this can lead to nonsensical results so you would use a logistic regression But at the end of the day we are just looking at the ratio of A, B, and C rather than the raw values we still have to do the regression to figure out which linear combination is the best. edit2: The->this | ||
|
Mylax
Germany21 Posts
You are saying that if I find a linear combination of KDA by using linear regression, this adjusted KDA will be inferior to a KDA that controls for the positive correlation between winning and kills, assists and negative correlation between winning and deaths? Another question I have is this: shouldn't the effect of winning on KDA and losing on KDA cancel each other out? | ||
|
LaNague
Germany9118 Posts
I waited 6 hours for the EUWest servers now. First they deactivate the games and say they prepare for a quick restart, then when the server restarts they say that it takes 2 hours and then after those 2 hours, they just dont say anything and assume that because its past midnight they can do whatever they want and noone will notice. So thats 6 hours downtime where they said 90 minutes. I guess it will be even more, lets see if the servers are up when i wake up. | ||
|
eagle
United States693 Posts
| ||
|
eagle
United States693 Posts
On August 22 2014 10:01 FakeSteve[TPR] wrote: I just read all of this thread from page 1 until now, haven't been around here much. There's something really significant in this thread that I want to address. ...They buffed MF so much that even shitcombo can win with her? rito? yeah hes 10 top miss forrune na now | ||
|
Scip
Czech Republic11293 Posts
On August 22 2014 10:01 FakeSteve[TPR] wrote: I just read all of this thread from page 1 until now, haven't been around here much. There's something really significant in this thread that I want to address. ...They buffed MF so much that even shitcombo can win with her? rito? Aren't there more pleasant way of killing your brain cells? A pickaxe maybe? | ||
|
Lmui
Canada6223 Posts
![[image loading]](http://1.bp.blogspot.com/-gd6InRM8khI/U_Z3aW7aIYI/AAAAAAAAV2o/hNFB7eXF6t0/s1600/Shyvana_Splash_0.jpg) ![[image loading]](http://1.bp.blogspot.com/-6TJZpPdViN8/U_Z3ab5GA7I/AAAAAAAAV2Y/bl6pJlmDc-E/s1600/Vayne_Splash_0.jpg) Links to the Shyvana/Vayne splashes. Shyv looks awesome, vayne looks very close to frostblade irelia.. | ||
|
Kaneh
Canada737 Posts
On August 21 2014 17:24 Goumindong wrote: IIRC that wouldn't help. There are no hubs in Texas. Well not quite true, but there aren't any hubs big enough for League. If there are a lot of servers in Texas its for tax and land cost reasons. (for instance WoW datacenters are in New York, Phoenix, LA, and Chicago) If you put something in Texas all the traffic will be routed from New York and LA anyway. (Probably routed from New York to LA to Texas). You won't gain any latency advantage from being in the middle of the US and you might even make it worse. People on the west coast have good connections not because of their physical proximity to the servers, but because LA and Seattle are the second and fourth(?) largest hubs in the US. Everyone on the west coast can go directly to LA or to Seattle and then directly to LA*. On the east coast everything goes through New York (IIRC) which means that if you aren't in New York you go there first, then you go to LA. If you live in Atlanta you go to from ALT to NY to LA. There could be more infrastructure now that allows them to skip the NY/LA jump. But that is how it used to be. *E.G. i sit basically on top of the Seattle hub. My ping to riot is about 12. Anyone who can get to Seattle relatively quickly (iirc there is also a hub in Portland, but not sure about that) can get to Riot about 12 ms later than that. Uhhhh. Houston has been a hub since forever and Dallas is growing so quickly it is a hub as well. Can you not talk about datacenters and Internet connectivity when you don't even know that Texas has the most datacenters outside of California? | ||
|
FinestHour
United States18466 Posts
| ||
|
Goumindong
United States3529 Posts
On August 22 2014 10:57 Mylax wrote: Ok, makes sense. So finding the linear combination and linear regression is the same. You are saying that if I find a linear combination of KDA by using linear regression, this adjusted KDA will be inferior to a KDA that controls for the positive correlation between winning and kills, assists and negative correlation between winning and deaths? Another question I have is this: shouldn't the effect of winning on KDA and losing on KDA cancel each other out? We don't have a way to know whether the effects cancel each other out or not(even if the effect was the same it would only cancel out for an individual player if their win rate was 50%). That is the problem, we can't just throw more data at the problem and get an answer when there is endogeneity. The only way we can(besides the 2 stage and a few other methods) is basically akin to knowing the function. But if you know the function then we don't need to do the stats to find it now do we? This is one of the reasons that Elo came up with ELO. Chessmasters evaluating the play of other chess players as they played was having hard time rating players. You couldn't say "he made more good looking plays" or "he took more pieces" or whatever. All that mattered was winning. The problem with individual champion ratings is that you can't quite just take someones MMR because they don't play all their games with that champion. If you knew the MMR of the games they played at and knew which ones they won and which ones they lost you could infer an individual champion MMR but we don't have that information without creating a massive database of games (like sufficiency has) and doing a lot of calculations for every player (which would be intensive) In the end, if you want to know if you do well with a champion just look at your win rate at your higher MMR's (basically ignore the stomps it took you to get from gold to diamond). Or simply, if you're positive with a champion at diamond 1 you're one of the best of that champion in the world | ||
| ||
