One example, the people who purposely feed. I never see these people. I had one unranked game with feeders. Ever. Other people see these people all the time. That isn't considered in elo in the slightest. You can argue that if you play enough games it will all even out. Except for two things. One, the amazing amount of variables not factored. Two, your elo changes based on these random factors which further skews things. Maybe if I play at least a few hundred games of ranked it will matter, but I don't play that much. i doubt a few hundred would be accurate anyway given other games can't get things right after thousands.
ELO is based on pretty well understood statistical and mathematical principles. Starcraft 2 has a pretty excellent system. It can place you decently well after 5 games. Pretty well after 20 games. And perfectly after 100 games (IIRC, it should be 95% confident of getting you within 20 elo of your 'true score' by 85 games. I'm a bit rusty on my stats but I think LoL requires 190 games.)
Keep in mind ELO isn't 'skill at the game'; it's a raw number based on 'winningness ability'. ELO is the Charlie Sheen coefficient. Feeders, AFKers, and disconnects are calculated into ELO. The game assumes that if you had more winningness ability, you would have convinced your team not to feed. (Ex, like when saintvicious went 4-0 by 6 minutes when he was in 'elo hell', none of his team gave up even though his jungler died to creeps and one member of the team was AFK.)
There's ample cases of people doing ELO challenges and raising themselves from sub-700 to 1400+, 1600+, or 2000+. The top 50 in all regions is full of alts. They never seem to run into these random problems on their either. And it doesn't take hundreds of games for them; normally around 100 suffices.
SC2 works because it is a 1v1 game like chess. LOL's Elo system is always going to be at the mercy of dumb fucking luck, because no one ever has a DC during a real life chess game, and moreover, people are less likely to troll a real life game because they realize there's a non-zero chance someone will break one or more bones (more is preferable).
Honestly, anyone who thinks bitching about LOL's elo is completely misguided is an idiot. Anyone who understands stats will agree that you'll eventually get to your "True elo," but the problem is that in a lot of cases, the journey there isn't fun. It only takes 2 trolls in a row, during placement, to lose you 100 Elo and sap all the fun out of your entire day's play time. Really, the problem is that the LOL community, barring most of the PA folks, sucks, and that Elo can give you an actual numerical value of how much they suck.
0
SpectrumArcher of InfernoChaldea Rec RoomRegistered Userregular
Corki is apparently some kind of counter to Vayne? I don't know, saw it a game or so ago, the guy playing him was drinking the Kool-aid and did indeed shit all over Vayne during laning.
Placement matches suck, but the alternative (having to play 30 games to get to 1500 instead of 8) is also pretty terrible for good players. Yeah, random chance has a huge effect, but the number of people who complain about Elo Hell is so high that it's easy to make the assumption that 90% of the Elo Hell players are really just either shitty players being Dunning-Krueger'd (sp?) to hell, or people who may be technically competent but have no clue what they should be doing to win ("I was forced to play support because my team sucks at picking, and I was dominating my lane in last hits using starcall, but our jungler never came to gank bot while their jungler got two double kills, it wasn't my fault!")
Corki is apparently some kind of counter to Vayne? I don't know, saw it a game or so ago, the guy playing him was drinking the Kool-aid and did indeed shit all over Vayne during laning.
vayne has a really shitty early game
corki can just walk up and Q her in the face in lane, and she can't really duel him early because he beats her
One thing I would really, REALLY like to know, is what really separates the different levels? What's the difference between 1100 and 1200? 1600 and 1700? I think the skill jump from 1100 to 1200 is more significant than 1600 to 1700, but I have no evidence.
It's hard to pick out any one thing, because people will rise in ELO for different reasons. Some people are really good mechanical players and dominate their lane. Others are good strategically and know how to make good calls on dragon/baron/buffs. Because of that, it's hard to pick out any one factor that leads to success in a specific ELO. For example, there was a 1100 player on the general forums the other day challenging 2K ELO players to 1v1 lanes and winning. A player like that might be very strong in one area relative to his ELO, but extremely weak in others very weak in others.
That said, I have noticed that each flavor. People in 1600-1700 tend to play hyper-aggressive, zoning even when it's unsafe to do so and blowing summoners on unconfirmed kills. By 1800-1900 people tend to play more passive and trust their teammates to punish aggression.
SC2 works because it is a 1v1 game like chess. LOL's Elo system is always going to be at the mercy of dumb fucking luck, because no one ever has a DC during a real life chess game, and moreover, people are less likely to troll a real life game because they realize there's a non-zero chance someone will break one or more bones (more is preferable).
Honestly, anyone who thinks bitching about LOL's elo is completely misguided is an idiot. Anyone who understands stats will agree that you'll eventually get to your "True elo," but the problem is that in a lot of cases, the journey there isn't fun. It only takes 2 trolls in a row, during placement, to lose you 100 Elo and sap all the fun out of your entire day's play time. Really, the problem is that the LOL community, barring most of the PA folks, sucks, and that Elo can give you an actual numerical value of how much they suck.
ELO works in SC2 because of dumb fucking luck too. It is in fact, the central assumption that makes ELO systems work.
It is indeed possible to have an ELO system misplace you, and the fewer games you play the more likely that is to happen. But is is very unlikely that you will stay there and in fact very likely that you will get to a reasonable level quickly.
That is, if you really are better than the rating the system gives you. It is much more the case that people simply are not better than their ELO, they cannot see it because they cannot see their own errors. Which is probably why they stay down there.
0
AriviaI Like A ChallengeEarth-1Registered Userregular
One example, the people who purposely feed. I never see these people. I had one unranked game with feeders. Ever. Other people see these people all the time. That isn't considered in elo in the slightest. You can argue that if you play enough games it will all even out. Except for two things. One, the amazing amount of variables not factored. Two, your elo changes based on these random factors which further skews things. Maybe if I play at least a few hundred games of ranked it will matter, but I don't play that much. i doubt a few hundred would be accurate anyway given other games can't get things right after thousands.
ELO is based on pretty well understood statistical and mathematical principles. Starcraft 2 has a pretty excellent system. It can place you decently well after 5 games. Pretty well after 20 games. And perfectly after 100 games (IIRC, it should be 95% confident of getting you within 20 elo of your 'true score' by 85 games. I'm a bit rusty on my stats but I think LoL requires 190 games.)
Keep in mind ELO isn't 'skill at the game'; it's a raw number based on 'winningness ability'. ELO is the Charlie Sheen coefficient. Feeders, AFKers, and disconnects are calculated into ELO. The game assumes that if you had more winningness ability, you would have convinced your team not to feed. (Ex, like when saintvicious went 4-0 by 6 minutes when he was in 'elo hell', none of his team gave up even though his jungler died to creeps and one member of the team was AFK.)
There's ample cases of people doing ELO challenges and raising themselves from sub-700 to 1400+, 1600+, or 2000+. The top 50 in all regions is full of alts. They never seem to run into these random problems on their either. And it doesn't take hundreds of games for them; normally around 100 suffices.
Managed to fend off Akali and Fizz for like 16 minutes (I had on AS/Armor runes on, too) while kaorti farmed with Tristana, I think only the enemy Graves out-farmed him due to being mid against Veigar. There was some hilarious lag, I randomly sauntered past two turrets and died to minions while Shyvana and Fizz ran into kaorti. He killed them both in an autoattack fight with like half his health left.
But seriously, Nidalee's heal on top of Tristana makes for some dead as fuck towers.
@AD Carries: The new Sivir is pretty baller and a half, on top of having some of the best skins. I prefer her to Vayne because Sivir can actually poke.
If anything, I suspect it's easier to luck into a higher ELO than a lower ELO. You can always pick support and get carried a few games in a row. If you're of significantly higher skill than the ELO you're in, it's pretty unlikely that you won't be able to carry a given game.
It doesn't. It is "place with x% confidence within a certain interval".
The more games you play, the smaller that interval gets, but the higher the % confidence the larger that interval gets.
However, all a 95% confidence interval means is "if we did this 100 times we would expect that 95% of the time the true value would lie within the generated interval"
Since there are more players in a game than you, each game contains less information about how good you are, so in order to produce high %confidence intervals with small margins you need to play more games.
Learn how to last hit and harass with q. Its not that hard. Learn map awareness for using your R.
You're going to be focused because of your AoE so when you die, try to die in advantageous positions, like right in front of your turret or right behind theirs.
r is a channel, so be careful about r'ing when you're alive if there are enemies with hard CC around.
Get lots of AP... press r... make pain
edit: and yea, ban soraka.
Edit: Core items are WotA, RoA, Deathcap, voidstaff. Not sure the order on that. RoA is survivability and sustain, WotA will make you spell vamp a LOT of HP with e(especially in a minion wave).
Also good is Zhonya's if it doesn't turn off your e.
Learn how to last hit and harass with q. Its not that hard. Learn map awareness for using your R.
You're going to be focused because of your AoE so when you die, try to die in advantageous positions, like right in front of your turret or right behind theirs.
r is a channel, so be careful about r'ing when you're alive if there are enemies with hard CC around.
Get lots of AP... press r... make pain
edit: and yea, ban soraka.
Edit: Core items are WotA, RoA, Deathcap, voidstaff. Not sure the order on that. RoA is survivability and sustain, WotA will make you spell vamp a LOT of HP with e(especially in a minion wave).
Also good is Zhonya's if it doesn't turn off your e.
Is tear of the goddess a good idea?
0
SpectrumArcher of InfernoChaldea Rec RoomRegistered Userregular
Learn how to last hit and harass with q. Its not that hard. Learn map awareness for using your R.
You're going to be focused because of your AoE so when you die, try to die in advantageous positions, like right in front of your turret or right behind theirs.
r is a channel, so be careful about r'ing when you're alive if there are enemies with hard CC around.
Get lots of AP... press r... make pain
edit: and yea, ban soraka.
Edit: Core items are WotA, RoA, Deathcap, voidstaff. Not sure the order on that. RoA is survivability and sustain, WotA will make you spell vamp a LOT of HP with e(especially in a minion wave).
Also good is Zhonya's if it doesn't turn off your e.
Is tear of the goddess a good idea?
It's pretty common on him, yeah
He stacks it really quickly and Karthus is always mana hungry
ELO is based on pretty well understood statistical and mathematical principles. Starcraft 2 has a pretty excellent system. It can place you decently well after 5 games. Pretty well after 20 games. And perfectly after 100 games (IIRC, it should be 95% confident of getting you within 20 elo of your 'true score' by 85 games. I'm a bit rusty on my stats but I think LoL requires 190 games.)
There's ample cases of people doing ELO challenges and raising themselves from sub-700 to 1400+, 1600+, or 2000+. The top 50 in all regions is full of alts. They never seem to run into these random problems on their either. And it doesn't take hundreds of games for them; normally around 100 suffices.
Why does LoL take 190 games to place you?
It's been a while since I did statistics, so if anyone is more up-to-date please chime in. Basically the weaker an effect has on the outcome, the more trials you need to be sure that it's having an effect. Like, if there was a cancer treatment that works 90% of the time, you'll be able to detect that it's doing something right away; however, if it only works 10% of the time, you'll need a lot more trials to be sure that it's the cancer treatment instead of you just getting lucky on patients.
In Starcraft, you're 1/2 of the contribution of success (the other 1/2 being your opponent). However you win or lose, be it because you disconnected, pulled off a great strategy, or floated your buidings around until the enemy leaves the game, you were 1/2 the reason for it. In League, you're expected to be 1/10th the reason for success when you reach ELO equilibrium. This means that there's more chance for statistical noise to effect the outcome of your matches, because you're 5 times less effective. However, being 5 times less effective doesn't mean you need 5 times more games to match it up; IIRC number of games needed increases logarithmically base 2 (If starcraft is .5 and LoL is .1, you need 2.24x as many games played in LoL to reach the same confidence interval). So, if we already know that Starcraft (or other 1v1 ELO systems) take 85 games to stabilize, then a 5v5 game system should stabilize at 190 games played.*
*LoL's and SC2's system of giving bonus weighting to your first games may shift this number a bit, but it should remain roughly accurate.
If anyone is more skilled with statistics please let me know if I've messed anything up. It has been a long time since I've used them in my daily life. I think it's right though.
Edit: And Goumindong said it better and more succinctly than I.
As always I'll reserve judging a champion till they've been out for a week or two but, man, the new champ sounds really boring.
Puppies: If you dislike the Eurolane so much just run a kill lane bottom. You won't always win but it will be more exciting.
0
SpectrumArcher of InfernoChaldea Rec RoomRegistered Userregular
The only issue with that is you can win your kill lane on bot, but enter midgame with no AD carry and then all of a sudden stalemates outside towers and teamfights are difficult.
The only issue with that is you can win your kill lane on bot, but enter midgame with no AD carry and then all of a sudden stalemates outside towers and teamfights are difficult.
Some AD carries (Corki jumps to mind) can solo top. But at that point, yeah, you are asking for a level of team comp coordination that may be quite hard to get with strangers.
0
SpectrumArcher of InfernoChaldea Rec RoomRegistered Userregular
The only issue with that is you can win your kill lane on bot, but enter midgame with no AD carry and then all of a sudden stalemates outside towers and teamfights are difficult.
Some AD carries (Corki jumps to mind) can solo top. But at that point, yeah, you are asking for a level of team comp coordination that may be quite hard to get with strangers.
Yep. Corki, Graves, Ez, and Vayne (possibly Cait) can all solo top. Easier to do in 5s than in solo, as you said.
i think getting rid of all burst healing in the game would be a small price to pay for the end of the eurolane
there are few things as boring to me to playing one of the supports in a relatively even eurolane
maybe playing ap sion is more boring
but that's all i can think of
A member of the dev team gave a very insightful post on the eurolane a while ago: the reason the eurolane exists isn't because of sustainy supports; it's because AD carries scale so well off of items that it is worth it to have their lanemate not take minion kills.
Interesting supports that you can play with a carry that aren't too boring are Blitzcrank, Alistar, and Leona.
Frost (Passive) - Sejuani's basic attacks apply Frost, reducing enemy Movement Speed by 10% for 3 seconds.
Arctic Assault - Sejuani charges forward to deal magic damage and apply Frost to enemies. Sejuani stops upon colliding with an enemy champion.
Northern Winds - Sejuani summons an arctic storm around her which deals magic damage to nearby enemies every second. Damage is increased against enemies affected by Frost or Permafrost.
Permafrost - Sejuani converts Frost on nearby enemies to Permafrost, dealing magic damage and increasing the Movement Speed reduction dramatically.
Glacial Prison (Ultimate) - Sejuani throws her weapon, stunning the first enemy champion hit. Nearby enemies are stunned for a shorter duration. All targets take magic damage and are affected by Frost.
I'm going to assume they are gonna give her Leona-like ratios. Seems like she can be pretty broken if she has decent ratios.
The ratios were apparently leaked, they're on the wiki. (Not sure if they're real ones, but the abilities are right, they've been posted at least two days ago.) http://leagueoflegends.wikia.com/wiki/Sejuani
Solid ratios but I don't really see building her as an ap carry, her cd's are just straight up too damn long. 11 (at max rank, 19 to start), 11, 10, 110 at max. Ap might be ok but you'd do tanky ap, like maokai.
LoL: failboattootoot
0
SpectrumArcher of InfernoChaldea Rec RoomRegistered Userregular
Well...Morello's+Ionian's(or FH) would fix alot of that.
The only issue with that is you can win your kill lane on bot, but enter midgame with no AD carry and then all of a sudden stalemates outside towers and teamfights are difficult.
Sona's an AD carry, what you talkin bout Spectrum. A carry with a heal and a movespeed buff! Sona/Taric the new meta in a few months time you'll see! You'll all see!
Sona's an AD carry, what you talkin bout Spectrum. A carry with a heal and a movespeed buff! Sona/Taric the new meta in a few months time you'll see! You'll all see!
Sona is totally an AD carry, yes. A carry with an enemy damage debuff and a AoE stun. Totally freaking OP
Sona winning 1v1s against an enemy carry due to timing W power chords well is the best thing
Sejuani sounds like everything that I hate about Ashe's slow toggle torn apart into 5 abilities and made into a champion more boring than Volibear.
Do not want.
Not even a little bit.
I would tend to say no. When your AoE is off you're going to get mana when you kill creeps and i think you're going to have plenty of base mana for a fight with the RoA. So i would tend to not get it unless you plan on building AA. It is pretty common on him, but i would think that a RoA is a better investment all around. You don't have to cast in order to get your sustain and with the AA fully stacked you're barely going to have enough mana to make the AA worth it(especially not compared to some other AP item for the cash, the RoA gives you HP and the catalyst sustain as well). On the other hand its really easy to stack mana because q is cheap and doesn't need a target to cast.
That being said, you want blue all the time and if you're getting it then buying an AA is a lot slower build into ability power than buying an RoA or WotA or Void Staff or Deathcap.
Re: AD's
Speaking of strange AD's why the fuck isn't Kennen a common AD carry...
575 range, 310 MS, high base health, good base AD, AS and scaling(tied for best base ranged attack speed in the game(teemo), second highest unbuffed attack speed at level 18(behind eve)). An on hit that will eventually stun and increases his DPS by 16% at level 5(+80% of attack damage every 5 attacks, so if you're counting and using the ability you can get a stun in 6 attacks) and grants more bonus damage, a chase that grants armor and mr, both of which can be used to proc on hit abilities and one of the best team fight ultimates in the game...
Edit: WRT: ELO and statistics
I am decent at stats but don't actually know anything of the inner workings of the ELO system used except the very basics of the theoretical construction and some intuition (basically it assumes your ELO is correct, assumes a distribution of possibilities over a normal range centered on that ELO and adjusts your new score based on what the prior belief of probability of winning was to a desired posterior [though i do not believe they do straight Bayesian updating]). I am not sure of the structure of confidence intervals but I can almost assure you that the number of games needed is not logarithmic since your posterior distribution is going to be some normal distribution(which is not going to have a closed form solution).
Now the actual calculation probably doesn't use the math below, it probably uses a simple algorithm which is based on the math below and ought to, as a random process, generate the math below. This, however, will make things complicated later, and once we relax a certain assumption will break what you consider the notion of a CI on your ELO.
To make this make a bit more sense, consider Bayes law [p(B|A) = p(A|B)p(B)/p(A). A,B are events and p(X|Y) is the probability of an event X given Y is true]. This works with distributions as well since we can write P(A|B) as a function of b. So if B is "your ELO" and A is "how well you did in the last game" what we can do is say, "Since we had a belief about how good you are P(B) we can multiply that by the probability we got the data we did (P(A|B), divide by the probability of winning in general and get a new distribution of your ELO"
Then we just solve that for your ELO. This is the basis of the ELO system, except that there are a few problems.
1) Calculating this is non-trivial. Well, maybe its trivial today, but i think that with the number of games going on, and the problems presented in writing a function that will make it all work, not to mention having to define a distribution for everyone rather than a simple rating, and having those distributions matter in the expectation function and worst of all, recalculating the entire thing over all your games every win due to another assumption not holding (specifically that your ELO can change)*.
2) The form of the distribution for winning has to be defined (and is probably defined in a logarithmic way based on the net difference in ELO). This has nothing to do with the standard deviations we might expect for your ELO since the central limit theorem holds. Defining this distribution is also probably non-trivial.
3) The big problem as hinted about above is that if we are using an algorithm we are not actually taking into account how wildly you play. So we don't actually have any information on what your particular CI might be, or anyone's really. We can assume standard deviations are the same for everyone and are fixed but we have no way of knowing whether or not these standard deviations are correct because we probably have no way of knowing anything about them**
Since we do know that there is less information recorded in a win or loss, we can say that compared to a 1v1 game the CI should be larger, but we don't have a way to actually generate it.
*If your true ELO can change then we have a problem in that all wins/losses are in the strict model weighted evenly. So if you lose 20 games and get better then those 20 games will effect your rating just as much as the last 20 you won. It would be like saying "Dan Dihn lost a game at 1200 elo which cancels out that 2000 win the other day" without nothing that that 1200 loss was from 2 years ago.
*To be fair, i don't know enough stats to know whether or not we can know anything about them. I have a feeling that whatever we try to observe is going to be flawed since in order to get your variance we're going to have to have everyone elses variances you've played with. And if we want to do that we're going to need your variance... This would be solvable if it did not result in a system of equations too large to solve.
IIRC number of games needed increases logarithmically base 2
No. That might be the expectation function (I.E. will you win or not, based on your ELO), but the normal distribution has no closed form and the distribution of your ELO will be normal due to the central limit theorem. The number of games needed to get to some specific %CI within a range of ELO will always be unknown.
OK: That being said, basically we can still be more sure of an ELO the more games you played, and so long as you've over a threshold.
Basically the justification goes like this:
1) At the start of a match, given everyones ELO and other conditionals, we assign each player an "expected value" for the match.(probably each team really) This is between zero and one and is logarithmic in nature.
Note: That the scale of the logarithm is chosen by the designers and will only serve to increase or decrease the standard deviation of the rating scale. IIRC chess uses log10 with an internal scaling of 400. I.E. every 400 points of ELO represents 10x better player.
2) At the end of the match each player gains a bonus or loss of points determined by a multiplier multiplied by the difference between the outcome and the expectation. Furthermore other calculations may be made in incorporate other information such as the last time you played.(See: Glicko, and the stuff that riot does to weight matches)
3) If we take this system and perform random experiments on the outcome assuming the distribution desired truly exists then the final ELO's will roughly match the normal distribution that we are expecting.
This is quite ideal because even if we do not believe that actual skill is organized in that way, it certainly is ordinal. However we can test to see (basically we check real people at the ratings and see if the win/loss in matches are what we would expect given the rating differentials, then we can check that back to see if the distribution we gave matches the actual distribution)
edit: Basically the process is of 3 is a) perform simulation based on known given distribution to check to make sure the algorithm produces that b) after the algorithm produces a distribution for real players, check to see if that distribution is right.
Goumindong on
0
Forever Zefirocloaked in the midnight glory of an event horizonRegistered Userregular
Everyone locked and support was left, but I've never played support in my life so I just picked Talon and went bottom with Sivir and instilled fear in our enemies
Supportalon won the game
by killing everyone
XBL - Foreverender | 3DS FC - 1418 6696 1012 | Steam ID | LoL
I am decent at stats but don't actually know anything of the inner workings of the ELO system used except the very basics of the theoretical construction and some intuition (basically it assumes your ELO is correct, assumes a distribution of possibilities over a normal range centered on that ELO and adjusts your new score based on what the prior belief of probability of winning was to a desired posterior [though i do not believe they do straight Bayesian updating]). I am not sure of the structure of confidence intervals but I can almost assure you that the number of games needed is not logarithmic since your posterior distribution is going to be some normal distribution(which is not going to have a closed form solution).
Now the actual calculation probably doesn't use the math below, it probably uses a simple algorithm which is based on the math below and ought to, as a random process, generate the math below. This, however, will make things complicated later, and once we relax a certain assumption will break what you consider the notion of a CI on your ELO.
To make this make a bit more sense, consider Bayes law [p(B|A) = p(A|B)p(B)/p(A). A,B are events and p(X|Y) is the probability of an event X given Y is true]. This works with distributions as well since we can write P(A|B) as a function of b. So if B is "your ELO" and A is "how well you did in the last game" what we can do is say, "Since we had a belief about how good you are P(B) we can multiply that by the probability we got the data we did (P(A|B), divide by the probability of winning in general and get a new distribution of your ELO"
Then we just solve that for your ELO. This is the basis of the ELO system, except that there are a few problems.
1) Calculating this is non-trivial. Well, maybe its trivial today, but i think that with the number of games going on, and the problems presented in writing a function that will make it all work, not to mention having to define a distribution for everyone rather than a simple rating, and having those distributions matter in the expectation function and worst of all, recalculating the entire thing over all your games every win due to another assumption not holding (specifically that your ELO can change)*.
2) The form of the distribution for winning has to be defined (and is probably defined in a logarithmic way based on the net difference in ELO). This has nothing to do with the standard deviations we might expect for your ELO since the central limit theorem holds. Defining this distribution is also probably non-trivial.
3) The big problem as hinted about above is that if we are using an algorithm we are not actually taking into account how wildly you play. So we don't actually have any information on what your particular CI might be, or anyone's really. We can assume standard deviations are the same for everyone and are fixed but we have no way of knowing whether or not these standard deviations are correct because we probably have no way of knowing anything about them**
Since we do know that there is less information recorded in a win or loss, we can say that compared to a 1v1 game the CI should be larger, but we don't have a way to actually generate it.
*If your true ELO can change then we have a problem in that all wins/losses are in the strict model weighted evenly. So if you lose 20 games and get better then those 20 games will effect your rating just as much as the last 20 you won. It would be like saying "Dan Dihn lost a game at 1200 elo which cancels out that 2000 win the other day" without nothing that that 1200 loss was from 2 years ago.
*To be fair, i don't know enough stats to know whether or not we can know anything about them. I have a feeling that whatever we try to observe is going to be flawed since in order to get your variance we're going to have to have everyone elses variances you've played with. And if we want to do that we're going to need your variance... This would be solvable if it did not result in a system of equations too large to solve.
IIRC number of games needed increases logarithmically base 2
No. That might be the expectation function (I.E. will you win or not, based on your ELO), but the normal distribution has no closed form and the distribution of your ELO will be normal due to the central limit theorem. The number of games needed to get to some specific %CI within a range of ELO will always be unknown.
OK: That being said, basically we can still be more sure of an ELO the more games you played, and so long as you've over a threshold.
Basically the justification goes like this:
1) At the start of a match, given everyones ELO and other conditionals, we assign each player an "expected value" for the match.(probably each team really) This is between zero and one and is logarithmic in nature.
Note: That the scale of the logarithm is chosen by the designers and will only serve to increase or decrease the standard deviation of the rating scale. IIRC chess uses log10 with an internal scaling of 400. I.E. every 400 points of ELO represents 10x better player.
2) At the end of the match each player gains a bonus or loss of points determined by a multiplier multiplied by the difference between the outcome and the expectation. Furthermore other calculations may be made in incorporate other information such as the last time you played.(See: Glicko, and the stuff that riot does to weight matches)
3) If we take this system and perform random experiments on the outcome assuming the distribution desired truly exists then the final ELO's will roughly match the normal distribution that we are expecting.
This is quite ideal because even if we do not believe that actual skill is organized in that way, it certainly is ordinal. However we can test to see (basically we check real people at the ratings and see if the win/loss in matches are what we would expect given the rating differentials, then we can check that back to see if the distribution we gave matches the actual distribution)
edit: Basically the process is of 3 is a) perform simulation based on known given distribution to check to make sure the algorithm produces that b) after the algorithm produces a distribution for real players, check to see if that distribution is right.
Lots of good stuff there. IIRC, Starcraft 2 indeed uses a bayesian "Elo" system. It comes from World of Warcraft's Arena system which had an extreme amount of reverse engineering done on it by tens of thousands of interested players. It uses a modified weighted system where games after a period of time will decay. Well, sorta; it's more complicated than that but that's basically a tracking system for moving skill*. Also, your writing correctly predicts that SC2 did indeed use an uncertainty subsystem and would track individual player performance variance. (I use quotes in "Elo" above. SC2 doesn't actually use the Elo system, but a variant like Glicko or Trueskill. Still they're all commonly called ELO and attempt to achieve the same goal anyhow.) Here's a link to the SC2 ELO thread.
*In WoW, paying for Arena Titles was a big deal, and so WoW's Arena system evolved a bunch of methods to combat inflating your skill by having ringers or swapping accounts. By the time I quit they were defeated too, but it's still probably the hardest to game team ELO system around.
Zileas, a LoL dev, mentioned that LoL uses a heavily modified custom designed ELO system, though I don't know if they specifically use Elo (as in, the one invented by Arpad Elo) or a Bayesian based ELO system. I think they use a simpler algorithm that's supposed to mimic bayesian updating in most of the cases, but is lighter on calculations. Still, even if it's not 100% perfect, I'd consider it a very accurate measurement of skill. Except for in a few edge cases, all the rating systems usually agree with each other decently enough.
zerg rush on
0
SpectrumArcher of InfernoChaldea Rec RoomRegistered Userregular
Speaking of strange AD's why the fuck isn't Kennen a common AD carry...
575 range, 310 MS, high base health, good base AD, AS and scaling(tied for best base ranged attack speed in the game(teemo), second highest unbuffed attack speed at level 18(behind eve)). An on hit that will eventually stun and increases his DPS by 16% at level 5(+80% of attack damage every 5 attacks, so if you're counting and using the ability you can get a stun in 6 attacks) and grants more bonus damage, a chase that grants armor and mr, both of which can be used to proc on hit abilities and one of the best team fight ultimates in the game...
Kennen has some AS/AD scaling but compared to a more traditional carry it's kinda eh. You can fit him into the same niche as AD Ahri; ie, sort of a thing if you're really good, but probably not what you want your team rely on.
Although, that's me being excessively gunshy over having fucking AS/AD Kennen going mid in the AP slot and having their ult doing jack and shit come midgame.
Posts
SC2 works because it is a 1v1 game like chess. LOL's Elo system is always going to be at the mercy of dumb fucking luck, because no one ever has a DC during a real life chess game, and moreover, people are less likely to troll a real life game because they realize there's a non-zero chance someone will break one or more bones (more is preferable).
Honestly, anyone who thinks bitching about LOL's elo is completely misguided is an idiot. Anyone who understands stats will agree that you'll eventually get to your "True elo," but the problem is that in a lot of cases, the journey there isn't fun. It only takes 2 trolls in a row, during placement, to lose you 100 Elo and sap all the fun out of your entire day's play time. Really, the problem is that the LOL community, barring most of the PA folks, sucks, and that Elo can give you an actual numerical value of how much they suck.
vayne has a really shitty early game
corki can just walk up and Q her in the face in lane, and she can't really duel him early because he beats her
tumblr | instagram | twitter | steam
It's hard to pick out any one thing, because people will rise in ELO for different reasons. Some people are really good mechanical players and dominate their lane. Others are good strategically and know how to make good calls on dragon/baron/buffs. Because of that, it's hard to pick out any one factor that leads to success in a specific ELO. For example, there was a 1100 player on the general forums the other day challenging 2K ELO players to 1v1 lanes and winning. A player like that might be very strong in one area relative to his ELO, but extremely weak in others very weak in others.
That said, I have noticed that each flavor. People in 1600-1700 tend to play hyper-aggressive, zoning even when it's unsafe to do so and blowing summoners on unconfirmed kills. By 1800-1900 people tend to play more passive and trust their teammates to punish aggression.
Replays here if someone wants to see pro Graves + Alistar action: Replay #1 feat. Support Pantheon and Replay #2 feat. Riot Lag Spikes
ELO works in SC2 because of dumb fucking luck too. It is in fact, the central assumption that makes ELO systems work.
It is indeed possible to have an ELO system misplace you, and the fewer games you play the more likely that is to happen. But is is very unlikely that you will stay there and in fact very likely that you will get to a reasonable level quickly.
That is, if you really are better than the rating the system gives you. It is much more the case that people simply are not better than their ELO, they cannot see it because they cannot see their own errors. Which is probably why they stay down there.
Why does LoL take 190 games to place you?
Managed to fend off Akali and Fizz for like 16 minutes (I had on AS/Armor runes on, too) while kaorti farmed with Tristana, I think only the enemy Graves out-farmed him due to being mid against Veigar. There was some hilarious lag, I randomly sauntered past two turrets and died to minions while Shyvana and Fizz ran into kaorti. He killed them both in an autoattack fight with like half his health left.
But seriously, Nidalee's heal on top of Tristana makes for some dead as fuck towers.
@AD Carries: The new Sivir is pretty baller and a half, on top of having some of the best skins. I prefer her to Vayne because Sivir can actually poke.
He's a strong champion if that matters to you.
Do you like to jungle? Because, while he can go top lane he is much better in the jungle
i bought karthus instead.
tips. tricks. fun builds?
It doesn't. It is "place with x% confidence within a certain interval".
The more games you play, the smaller that interval gets, but the higher the % confidence the larger that interval gets.
However, all a 95% confidence interval means is "if we did this 100 times we would expect that 95% of the time the true value would lie within the generated interval"
Since there are more players in a game than you, each game contains less information about how good you are, so in order to produce high %confidence intervals with small margins you need to play more games.
there are few things as boring to me to playing one of the supports in a relatively even eurolane
maybe playing ap sion is more boring
but that's all i can think of
press R
tumblr | instagram | twitter | steam
And ban Soraka if you can.
Learn how to last hit and harass with q. Its not that hard. Learn map awareness for using your R.
You're going to be focused because of your AoE so when you die, try to die in advantageous positions, like right in front of your turret or right behind theirs.
r is a channel, so be careful about r'ing when you're alive if there are enemies with hard CC around.
Get lots of AP... press r... make pain
edit: and yea, ban soraka.
Edit: Core items are WotA, RoA, Deathcap, voidstaff. Not sure the order on that. RoA is survivability and sustain, WotA will make you spell vamp a LOT of HP with e(especially in a minion wave).
Also good is Zhonya's if it doesn't turn off your e.
Is tear of the goddess a good idea?
He stacks it really quickly and Karthus is always mana hungry
It's been a while since I did statistics, so if anyone is more up-to-date please chime in. Basically the weaker an effect has on the outcome, the more trials you need to be sure that it's having an effect. Like, if there was a cancer treatment that works 90% of the time, you'll be able to detect that it's doing something right away; however, if it only works 10% of the time, you'll need a lot more trials to be sure that it's the cancer treatment instead of you just getting lucky on patients.
In Starcraft, you're 1/2 of the contribution of success (the other 1/2 being your opponent). However you win or lose, be it because you disconnected, pulled off a great strategy, or floated your buidings around until the enemy leaves the game, you were 1/2 the reason for it. In League, you're expected to be 1/10th the reason for success when you reach ELO equilibrium. This means that there's more chance for statistical noise to effect the outcome of your matches, because you're 5 times less effective. However, being 5 times less effective doesn't mean you need 5 times more games to match it up; IIRC number of games needed increases logarithmically base 2 (If starcraft is .5 and LoL is .1, you need 2.24x as many games played in LoL to reach the same confidence interval). So, if we already know that Starcraft (or other 1v1 ELO systems) take 85 games to stabilize, then a 5v5 game system should stabilize at 190 games played.*
If anyone is more skilled with statistics please let me know if I've messed anything up. It has been a long time since I've used them in my daily life. I think it's right though.
Edit: And Goumindong said it better and more succinctly than I.
Puppies: If you dislike the Eurolane so much just run a kill lane bottom. You won't always win but it will be more exciting.
Some AD carries (Corki jumps to mind) can solo top. But at that point, yeah, you are asking for a level of team comp coordination that may be quite hard to get with strangers.
A member of the dev team gave a very insightful post on the eurolane a while ago: the reason the eurolane exists isn't because of sustainy supports; it's because AD carries scale so well off of items that it is worth it to have their lanemate not take minion kills.
Interesting supports that you can play with a carry that aren't too boring are Blitzcrank, Alistar, and Leona.
Handmade Jewelry by me on EtsyGames for sale
Me on Twitch!
Solid ratios but I don't really see building her as an ap carry, her cd's are just straight up too damn long. 11 (at max rank, 19 to start), 11, 10, 110 at max. Ap might be ok but you'd do tanky ap, like maokai.
LoL: failboattootoot
Jungle kayle into AD carry kayle
do eet
3ds friend code: 2981-6032-4118
Handmade Jewelry by me on EtsyGames for sale
Me on Twitch!
I recently tried going back to Alistar after playing so many Taric games and he just doesn't cut it. Taric does so much more burst damage.
Sona winning 1v1s against an enemy carry due to timing W power chords well is the best thing
Do not want.
Not even a little bit.
I would tend to say no. When your AoE is off you're going to get mana when you kill creeps and i think you're going to have plenty of base mana for a fight with the RoA. So i would tend to not get it unless you plan on building AA. It is pretty common on him, but i would think that a RoA is a better investment all around. You don't have to cast in order to get your sustain and with the AA fully stacked you're barely going to have enough mana to make the AA worth it(especially not compared to some other AP item for the cash, the RoA gives you HP and the catalyst sustain as well). On the other hand its really easy to stack mana because q is cheap and doesn't need a target to cast.
That being said, you want blue all the time and if you're getting it then buying an AA is a lot slower build into ability power than buying an RoA or WotA or Void Staff or Deathcap.
Re: AD's
Speaking of strange AD's why the fuck isn't Kennen a common AD carry...
575 range, 310 MS, high base health, good base AD, AS and scaling(tied for best base ranged attack speed in the game(teemo), second highest unbuffed attack speed at level 18(behind eve)). An on hit that will eventually stun and increases his DPS by 16% at level 5(+80% of attack damage every 5 attacks, so if you're counting and using the ability you can get a stun in 6 attacks) and grants more bonus damage, a chase that grants armor and mr, both of which can be used to proc on hit abilities and one of the best team fight ultimates in the game...
Edit: WRT: ELO and statistics
Now the actual calculation probably doesn't use the math below, it probably uses a simple algorithm which is based on the math below and ought to, as a random process, generate the math below. This, however, will make things complicated later, and once we relax a certain assumption will break what you consider the notion of a CI on your ELO.
To make this make a bit more sense, consider Bayes law [p(B|A) = p(A|B)p(B)/p(A). A,B are events and p(X|Y) is the probability of an event X given Y is true]. This works with distributions as well since we can write P(A|B) as a function of b. So if B is "your ELO" and A is "how well you did in the last game" what we can do is say, "Since we had a belief about how good you are P(B) we can multiply that by the probability we got the data we did (P(A|B), divide by the probability of winning in general and get a new distribution of your ELO"
Then we just solve that for your ELO. This is the basis of the ELO system, except that there are a few problems.
1) Calculating this is non-trivial. Well, maybe its trivial today, but i think that with the number of games going on, and the problems presented in writing a function that will make it all work, not to mention having to define a distribution for everyone rather than a simple rating, and having those distributions matter in the expectation function and worst of all, recalculating the entire thing over all your games every win due to another assumption not holding (specifically that your ELO can change)*.
2) The form of the distribution for winning has to be defined (and is probably defined in a logarithmic way based on the net difference in ELO). This has nothing to do with the standard deviations we might expect for your ELO since the central limit theorem holds. Defining this distribution is also probably non-trivial.
3) The big problem as hinted about above is that if we are using an algorithm we are not actually taking into account how wildly you play. So we don't actually have any information on what your particular CI might be, or anyone's really. We can assume standard deviations are the same for everyone and are fixed but we have no way of knowing whether or not these standard deviations are correct because we probably have no way of knowing anything about them**
Since we do know that there is less information recorded in a win or loss, we can say that compared to a 1v1 game the CI should be larger, but we don't have a way to actually generate it.
*If your true ELO can change then we have a problem in that all wins/losses are in the strict model weighted evenly. So if you lose 20 games and get better then those 20 games will effect your rating just as much as the last 20 you won. It would be like saying "Dan Dihn lost a game at 1200 elo which cancels out that 2000 win the other day" without nothing that that 1200 loss was from 2 years ago.
*To be fair, i don't know enough stats to know whether or not we can know anything about them. I have a feeling that whatever we try to observe is going to be flawed since in order to get your variance we're going to have to have everyone elses variances you've played with. And if we want to do that we're going to need your variance... This would be solvable if it did not result in a system of equations too large to solve.
No. That might be the expectation function (I.E. will you win or not, based on your ELO), but the normal distribution has no closed form and the distribution of your ELO will be normal due to the central limit theorem. The number of games needed to get to some specific %CI within a range of ELO will always be unknown.
OK: That being said, basically we can still be more sure of an ELO the more games you played, and so long as you've over a threshold.
Basically the justification goes like this:
1) At the start of a match, given everyones ELO and other conditionals, we assign each player an "expected value" for the match.(probably each team really) This is between zero and one and is logarithmic in nature.
Note: That the scale of the logarithm is chosen by the designers and will only serve to increase or decrease the standard deviation of the rating scale. IIRC chess uses log10 with an internal scaling of 400. I.E. every 400 points of ELO represents 10x better player.
2) At the end of the match each player gains a bonus or loss of points determined by a multiplier multiplied by the difference between the outcome and the expectation. Furthermore other calculations may be made in incorporate other information such as the last time you played.(See: Glicko, and the stuff that riot does to weight matches)
3) If we take this system and perform random experiments on the outcome assuming the distribution desired truly exists then the final ELO's will roughly match the normal distribution that we are expecting.
This is quite ideal because even if we do not believe that actual skill is organized in that way, it certainly is ordinal. However we can test to see (basically we check real people at the ratings and see if the win/loss in matches are what we would expect given the rating differentials, then we can check that back to see if the distribution we gave matches the actual distribution)
edit: Basically the process is of 3 is a) perform simulation based on known given distribution to check to make sure the algorithm produces that b) after the algorithm produces a distribution for real players, check to see if that distribution is right.
Supportalon won the game
by killing everyone
XBL - Foreverender | 3DS FC - 1418 6696 1012 | Steam ID | LoL
Lots of good stuff there. IIRC, Starcraft 2 indeed uses a bayesian "Elo" system. It comes from World of Warcraft's Arena system which had an extreme amount of reverse engineering done on it by tens of thousands of interested players. It uses a modified weighted system where games after a period of time will decay. Well, sorta; it's more complicated than that but that's basically a tracking system for moving skill*. Also, your writing correctly predicts that SC2 did indeed use an uncertainty subsystem and would track individual player performance variance. (I use quotes in "Elo" above. SC2 doesn't actually use the Elo system, but a variant like Glicko or Trueskill. Still they're all commonly called ELO and attempt to achieve the same goal anyhow.) Here's a link to the SC2 ELO thread.
*In WoW, paying for Arena Titles was a big deal, and so WoW's Arena system evolved a bunch of methods to combat inflating your skill by having ringers or swapping accounts. By the time I quit they were defeated too, but it's still probably the hardest to game team ELO system around.
Zileas, a LoL dev, mentioned that LoL uses a heavily modified custom designed ELO system, though I don't know if they specifically use Elo (as in, the one invented by Arpad Elo) or a Bayesian based ELO system. I think they use a simpler algorithm that's supposed to mimic bayesian updating in most of the cases, but is lighter on calculations. Still, even if it's not 100% perfect, I'd consider it a very accurate measurement of skill. Except for in a few edge cases, all the rating systems usually agree with each other decently enough.
Although, that's me being excessively gunshy over having fucking AS/AD Kennen going mid in the AP slot and having their ult doing jack and shit come midgame.
They don't always mention all the changes that will be in the patch notes.