| Forum Index > LoL General |
|
cLutZ
United States19573 Posts
On December 08 2014 12:31 chalice wrote: yeah, korean solo queue talent has basically never been able translate into competitive success. You're supposed to put /s at the end. | ||
|
Dark_Chill
Canada3353 Posts
Edit: all of these should be team differences. | ||
|
chalice
United States1945 Posts
On December 08 2014 12:32 red_ wrote: You're basically resigned to making up stats and looking for correlations until you find one that maybe sticks. you say this like it isn't basically the foundation of every scientific research work done ever. all i'm trying to do is see if anyone has any interesting hypotheses about what makes one player more skilled/win more games than another that might be worth testing. On December 08 2014 13:01 Dark_Chill wrote: Well, that's the thing. You don't say that kills and deaths cause wins and losses. But you can say that high kills and low deaths can contribute to wins and losses. I think that going for teamwide stats first are better choice. Kill differential between the two teams might be a good stat. Dragons and Barons should be pretty good as well. I don't know if it's possible to calculate this, but an integral of map coverage (percent of map in vision) over the entire game. objective and vision control are definitely going to be a big part of my early focus. | ||
|
Gahlo
United States35117 Posts
On December 08 2014 13:05 chalice wrote: you say this like it isn't basically the foundation of every scientific research work done ever. all i'm trying to do is see if anyone has any interesting hypotheses about what makes one player more skilled/win more games than another that might be worth testing. That's simple: talent, mindset, and communication. | ||
|
Dark_Chill
Canada3353 Posts
On December 08 2014 13:07 Gahlo wrote: That's simple: talent, mindset, and communication. I'm pretty sure you're just trying to make fun of him now. Unless you're trying to say that those are stats that are available in the game. | ||
|
red_
United States8474 Posts
On December 08 2014 13:05 chalice wrote: you say this like it isn't basically the foundation of every scientific research work done ever. all i'm trying to do is see if anyone has any interesting hypotheses about what makes one player more skilled/win more games than another that might be worth testing. objective and vision control are definitely going to be a big part of my early focus. There's generally a more concrete starting point though. Like, using the baseball example, we've always known that good hitters tend to you know, hit the ball, and contribute to scoring runs. What is the league equivalent? CS? Well that's dependent on a billion things, and isn't even a goal at all for some champions/roles. Is the game involving a lot of grouping and team roaming? That's going to heavily impact numbers. Lane swaps? You're a support? A jungler that basically stops farming after X item progression? Then you have to consider that sports have had relatively minor rule changes in their history. An 'era' in sports generally lasts decades. We talk about basketball statistics in 'the shot clock era' not 'patch 4.19.' It's a nightmare for consistent data measurement, which is immensely important to advanced metrics. | ||
|
Goumindong
United States3529 Posts
On December 08 2014 13:05 chalice wrote: you say this like it isn't basically the foundation of every scientific research work done ever. all i'm trying to do is see if anyone has any interesting hypotheses about what makes one player more skilled/win more games than another that might be worth testing. No. The foundation of scientific research is starting with a model that fits a reasonable theory then testing that model only. The reason this is done is because hunting for correlations will get you correlations with certainty, but they will have high likelihoods of being essentially false. The thing is, we know what makes players better, mechanics, decision making, positioning etc. What we don't have is a way of measuring those values in a way that means anything. This is partially because we don't track everything and partially because the structure of the game presents statistical issues which are hard to overcome. So if you want to predict the winner of a soloqueue game then "weighted MMR at role" is what you're looking for. And if you're trying to find diamonds in the rough in soloqueue then "MMR at role" is what you're looking for. | ||
|
chalice
United States1945 Posts
On December 08 2014 13:38 Goumindong wrote: No. The foundation of scientific research is starting with a model that fits a reasonable theory then testing that model only. The reason this is done is because hunting for correlations will get you correlations with certainty, but they will have high likelihoods of being essentially false. The thing is, we know what makes players better, mechanics, decision making, positioning etc. What we don't have is a way of measuring those values in a way that means anything. This is partially because we don't track everything and partially because the structure of the game presents statistical issues which are hard to overcome. So if you want to predict the winner of a soloqueue game then "weighted MMR at role" is what you're looking for. And if you're trying to find diamonds in the rough in soloqueue then "MMR at role" is what you're looking for. lol, how do you not understand that sorting by MMR alone is completely worthless for finding undervalued talent when it's the only measurement of skill that everyone agrees on? | ||
|
Gahlo
United States35117 Posts
On December 08 2014 14:01 chalice wrote: lol, how do you not understand that sorting by MMR alone is completely worthless for finding undervalued talent when it's the only measurement of skill that everyone agrees on? On December 08 2014 13:38 Goumindong wrote: So if you want to predict the winner of a soloqueue game then "weighted MMR at role" is what you're looking for. And if you're trying to find diamonds in the rough in soloqueue then "MMR at role" is what you're looking for. Real? | ||
|
ShaLLoW[baY]
Canada12499 Posts
| ||
|
Sufficiency
Canada23833 Posts
On December 08 2014 08:02 chalice wrote: i'm correct in assuming that the people with the "all stats are useless" opinion have made virtually zero effort to actually analyze any data, right? i think observing that stats like KDA and CS are severely lacking when it comes to predicting the probability of victory and coming to the conclusion that all stats are worthless is a closed-minded and very limited way of thinking. just because the current, readily available stats suck doesn't mean that better metrics can't be developed. maybe i asked in the wrong way, but i was hoping to get some insights from higher elo TLers about what aspects of the game they think are important and improve your chances to win the match. Actually, KDA is an incredible predictor of win rates (red = game won; black = game lost). ![[image loading]](http://i.imgur.com/6l91qrc.png) I eyeballed a predictor of ln (1 + KDA) >= 1. This gave me a mean prediction error of 0.22 on my validation set. Roughly speaking, by looking at a 100k+ sample of gold/Plat games in patch 4.17, I can predict wins by checking KDA >= 1.7 and be correct 78% of the time. That is very impressive by itself. Remind you that I merely "eyeballed" this. The actual predictive power is far larger than what I described. | ||
|
chalice
United States1945 Posts
On December 08 2014 14:11 ShaLLoW[baY] wrote: i have literally no idea what the point of some of these arguments is besides just arguing for argument's sake yeah, it's okay if no one has any interesting ideas to share and i appreciate the concern for how i spend my time, but shockingly enough, my world is not going to end if i fail to develop an accurate predictive model with real world practical use based on my analysis of video game performance data. developing skills and gaining experience doing actual analysis of data that i find interesting is valuable enough on its own that doing the research for research's sake is plenty worthwhile. | ||
|
GhandiEAGLE
United States20754 Posts
| ||
|
Goumindong
United States3529 Posts
On December 08 2014 14:17 Sufficiency wrote: Actually, KDA is an incredible predictor of win rates (red = game won; black = game lost). I eyeballed a predictor of ln (1 + KDA) >= 1. This gave me a mean prediction error of 0.22 on my validation set. Roughly speaking, by looking at a 100k+ sample of gold/Plat games in patch 4.17, I can predict wins by checking KDA >= 1.7 and be correct 78% of the time. That is very impressive by itself. Remind you that I merely "eyeballed" this. The actual predictive power is far larger than what I described. Reverse causation, son. We aren't trying to predict a games in which we know the KDA in that game. After all, we also know the outcome so who cares? We are trying to predict the games for which we know some sort of "prior KDA" and the answer to that is "that shit don't work use MMR" On December 08 2014 14:01 chalice wrote: lol, how do you not understand that sorting by MMR alone is completely worthless for finding undervalued talent when it's the only measurement of skill that everyone agrees on? LOL how do you not understand that there exists no other way to find diamonds in the rough because no such measurement exists or can exist? Literally you're asking "I want to know how good people are at winning" and I am sitting here saying "this is the stat which shows how good people are at winning, its literally the value for their winningness weighted against competition" | ||
|
Gahlo
United States35117 Posts
On December 08 2014 14:38 GhandiEAGLE wrote: I see you, Mr. Black Dot up there at the top. Someone's in ELO hell. Sup. | ||
|
Sufficiency
Canada23833 Posts
On December 08 2014 14:44 Goumindong wrote: Reverse causation, son. We aren't trying to predict a games in which we know the KDA in that game. After all, we also know the outcome so who cares? We are trying to predict the games for which we know some sort of "prior KDA" and the answer to that is "that shit don't work use MMR" I don't think you know what you are talking about and you should stop further embarrassing yourself with the pretentious know-it-all act. No one cares about causation here. | ||
|
Goumindong
United States3529 Posts
On December 08 2014 14:46 Sufficiency wrote: I don't think you know what you are talking about and you should stop further embarrassing yourself with the pretentious know-it-all act. No one cares about causation here. Yes, we do. Chalice is asking for a way in which to identify soloqueue players who could totally go pro. That isn't "look at the stats in a single game and then predict that single game" both because that is trivial, but also because that doesn't actually tell us how good a player on the team is, so we can't use that to determine whether or not they're a diamond in the rough. KDA has reverse causation issues. In that winning causes higher KDA's(teams that are behind must take more risks in order to get back in the game, and so die more). Such we can't just look at people with high KDA's and think that it means anything, certainly it doesn't mean anything more than MMR when looking at how good a particular player is. Its fine if you want to predict the game in which the KDA's occurred. But then we're back to "predicting a game based on the events of the game is trivial". | ||
|
Ryuu314
United States12679 Posts
| ||
|
Gahlo
United States35117 Posts
On December 08 2014 14:54 Goumindong wrote: Yes, we do. Chalice is asking for a way in which to identify soloqueue players who could totally go pro. That isn't "look at the stats in a single game and then predict that single game" both because that is trivial, but also because that doesn't actually tell us how good a player on the team is, so we can't use that to determine whether or not they're a diamond in the rough. KDA has reverse causation issues. In that winning causes higher KDA's(teams that are behind must take more risks in order to get back in the game, and so die more). Such we can't just look at people with high KDA's and think that it means anything, certainly it doesn't mean anything more than MMR when looking at how good a particular player is. Its fine if you want to predict the game in which the KDA's occurred. But then we're back to "predicting a game based on the events of the game is trivial". Didn't we already solve this discussion with positional MMR? | ||
|
Goumindong
United States3529 Posts
| ||
| ||
|
|
Replay Cast
GSL Code S
Maru vs TriGGeR
Rogue vs NightMare
The PondCast
Replay Cast
OSC
Online Event
SOOP
CranKy Ducklings
WardiTV Invitational
SC Evo League
[ Show More ] |
|
|
EPT
