Ladder Decay

[Stratos]

I don't know if other ladders have this problem but the doubles ladder is very sparse at the top (this is everyone within 200 points of number one), which can make the current ladder decay algorithm a pain in the ass. I'm not 100% sure how the current decay algorithm works but I know it takes points off of your true ELO and even though i play a ladder game every day I lost 5 points off of my ELO every time i woke up for these past couple of days.

Now because so few people actually give me more than 6-7 points and because I don't want to, or don't have time to, play many ladder games per day (finding a battle on Thursday took me forty minutes) I'm basically just treading water ELO-wise, despite the fact that I haven't stopped winning battles. On more active / less bottom-heavy ladders these problems are much less apparent, but the question still remains of why an active ladderer who only wins games isn't seeing their ELO score increase or, honestly, why the decay system should detract from your "true score" at all?

Obviously I'm in favor of inactivity affecting your rating so people can't just get a lucky streak and then "park" an alt on the top and stop playing to avoid losing points to a loss. But I think there's a much more elegant way to do this, that doesn't punish people who aren't doing this. And if the people who "parked" come back after a while and prove they've still got it, I don't think that they deserve to permanently lose the points that they rightfully earned back before they took a hiatus for whatever reason.

I remember back in 2011 or so I read a post from PO about how they did their decay system. Basically there was your true ELO, which wasn't touched, and then it was adjusted by a function of your activity whenever it was displayed. The equation was simply:

D = Displayed ELO
T = True ELO
b = exponential base, <1
x = inactivity tracker
t = inactivity period

For every period of t in which no games are played, x = x+1. For every game played, x = x-1 (floor zero).
D = T*b^x.

So for example if t = 2 days and b = 0.95: If I don't play for four days, my inactivity tracker will be 2. My displayed ELO will be my true ELO times 0.95^2. If I then play a game, my true ELO will be adjusted by however many points were won/lost, but also x will drop to 1, which will increase my displayed ELO. In that way, long streaks of inactivity are "forgiven" when you become active again, rather than having to ladder simply to earn back points you lost by not playing. I think this is a much better method of handling inactivity.

i know this explanation sucks so if u want something clarified just ask

 

[Zarel]

So, without permanent ladder decay, your score will just keep on going up and up (i.e. rating inflation). That seems kind of bad for a ladder, doesn't it? A high number should mean you're a good player, not a player who's played for a long time, right?

It seems like your score shouldn't be increasing if you're not getting better. That's what a ladder rating is, isn't it? An estimate of your skill?

 

[dice]

w. po's rating decay they stored your true elo, and the 'decayed' amount was regained/lost to a certain degree depending on whether you won or lost your next few games (iirc?). it's a system that still punishes you for not playing, but you don't permanently lose the points as well. p sure pwnemon just thinks this compromise is superior to a flat-out decrease.

 

[Zarel]

w. po's rating decay they stored your true elo, and the 'decayed' amount was regained/lost to a certain degree depending on whether you won or lost your next few games (iirc?). it's a system that still punishes you for not playing, but you don't permanently lose the points as well. p sure pwnemon just thinks this compromise is superior to a flat-out decrease.

ok, but that has nothing to do with my post, you're just repeating what Stratos said

 

[dice]

the points lost via the decay aren't immediately accessible and are dependent on the outgame of more games.

So, without permanent ladder decay, your score will just keep on going up and up (i.e. rating inflation).

if you win games, sure. wouldn't this be true regardless of the system proposed if they played on a daily basis?

i was reiterating his post because your response seemed to tangentially tackle it. this system is a midway point between permanent and impermanent decay; it forces players to have to play and win games to rise the ladder while not harshly imposing a huge loss on players on vacation from the game. the 'decayed' elo drops alongside your actual elo if you lose a game, so you're effectively losing twice the elo per game. still a punishment, just not as severe.

 

[marilli]

I think someone didn't really get what zarel meant by
"It seems like your score shouldn't be increasing if you're not getting better. That's what a ladder rating is, isn't it? An estimate of your skill?"

As a player I know where people are coming from. When people play games on the ladder, they want to be rewarded with points. This quick and instant reward system is great feelings for the player and the best way to get you hooked and keep playing. This is good for the ranking system in a way, because it encourages good players to play more. But let's stop thinking from the players' perspective for just a sec, and instead from the perspective of the ladder, trying to figure out who's good and who's bad. The decay not exists because "yeah that lucky streaker who's parking that alt there, he shouldn't be there, that's gay." It does because after not playing for a while the system expects you to revert to the norm. That's why when you are a below norm player you don't decay. (probably also lessens the computation load by not calculating stuff for heaps of 1100 elo players, but that's not the main point, the point is that there's a good justification for it)

Every game is a data point trying to figure out how good you are. If your winrate vs ladder noobs is expected to be super high (like Stratos' case), you'll probably get really close to the winrate as the ladder expects. Why? Because Pokemon is ultimately a game of luck. So the ladder sees that your winrate vs ladder noobs is very well estimated by your score differential vs. them. So you're just treading water. It dindu nuffin wrong, it's perfectly working as intended. What's wrong with it, and why should you gain points?, ladder asks.

Maybe your point is that the decay is too large for a game that you really don't lose large amounts of skill over just because you haven't played recently, or something like that. That could be a valid point of contention, though it is not my place nor my authority. Maybe that can be considered by Zarel though I really do think he knows for the best, and probably still leads to inflation which is an issue. The suggestion of decay not applying to your true ELO simply isn't a good idea. It's not a "compromise" as Dice put it. The thought that this is some sort of a compromise - a middle ground - obviously comes from the idea that it's the middle ground of accounting for activity and letting people have fun by gaining points. However, only thing it's compromising is the ranking system, at cost of player experience.

This is essentially the difficulty in pleasing the playerbase, preventing inflation, trying to estimate people's true skills in a luck based game, etc. all at once. Every mathematical model has flaws, but this is just particularly worse. You just can't do it all. Zarel has made administrative choice already and I see no valid reason for him to change his opinion. I'm actually just mirroring Zarel's first post, and just blew it up into however long this is. I think the user feedback that it's frustrating as heck as a player is a relevant one, but I don't think Zarel's being unreasonably stubborn or anything here.

 

[Stratos]

So, without permanent ladder decay, your score will just keep on going up and up (i.e. rating inflation). That seems kind of bad for a ladder, doesn't it? A high number should mean you're a good player, not a player who's played for a long time, right?

It seems like your score shouldn't be increasing if you're not getting better. That's what a ladder rating is, isn't it? An estimate of your skill?

my score will go down if i lose. If I win two hundred games in a row, why shouldn't my score be in the exosphere? Every game I play I'll be looking at a +1/-39 spread and if i can keep my rating going up despite that by maintaining at least a 40:1 W/L ratio versus players near the top of the ladder, I think I've earned it. And if i try to "park" because I don't think I'll be able to maintain my ladder score, then it will start decaying and I'll be forced to play and risk losing 39 points. Your score doesn't "just keep going up and up" unless you do not lose; that's the point of ELO, to eventually find a rough equilibrium point where your wins and losses multiplied by their weights will balance out and keep you looping around in roughly the same skill level relative area. The only thing we need ladder decay for is to force people to keep playing so that ELO can do its job.

 

[Antar]

Rating inflation is actually a technical problem with Elo, and Zarel's implementation specifically (one that I think is well resolved by having actual decay).

This is a bit technical, but here goes:

• Elo as originally defined is "conservative" in that the net number of points is preserved--one player gains only as much as the other loses. This means that the average rating is always the starting rating (in our case 1000). However, even with "vanilla Elo," rating inflation is still a problem. First, new accounts mean there's consistently new points "entering the system." And the players (alts) that leave tend to have below-average ratings. So if you look at the average rating among ACTIVE players, even with vanilla Elo, rating inflation will still happen.
• PS does not use vanilla Elo. Specifically, IIRC, if your rating is above a certain point, you gain/lose fewer points than if your rating is below a certain point. This should in theory slow down rating inflation (higher-rated players will have difficulty climbing higher) but it also means that they're less likely to lose points.
• PS also, significantly, does not allow a player's Elo to drop below 1000 (this was a wise decision done to encourage players to NOT alt-reset. But this means that ladder inflation is an even more significant problem.
• Decay is the MOST STRAIGHTFORWARD WAY to deal with this problem. And it NEEDS to be an actual decay of rating, NOT just a display-thing like it was for PO. Stratos, I'm not saying your concern isn't valid, but the solution is adjust the decay algorithm not to ditch it.

My personal feeling is that Elo (ZElo especially) is likely a poor skill estimator (I'm going to put it on my to-do list to numerically analyze this), and it's why I feel we should be using GXE to rank people at the tops of the ladders. But I've been fighting that fight for a long time, and I'm no closer to winning it now than I was.

Stratos, now let me dissect your complaint:

Now because so few people actually give me more than 6-7 points

This is how Elo works. The higher your rating vs. your opponent, the fewer points you gain by winning. Because it's SUPPOSED to be a skill-estimator and because we DON'T want people at the top being there simply because they've played for longer. If you don't like that in a ladder (we had a LONG discussion about this a few years ago), you probably want a different ladder score system in place (I invented one called ARMS that would have suited you).

(finding a battle on Thursday took me forty minutes)

This is a serious concern and means Zarel needs to take another look at the matchmaking algorithm.

I'm basically just treading water ELO-wise, despite the fact that I haven't stopped winning battles.

Are you a significantly better player today than you were last week? Ratings are SUPPOSED to stabilize. They're supposed to reflect your skill, NOT your accomplishment. I wrote a big long guide about this. Again, if you want the ladder to measure accomplishment, you want a different ladder system, and one that isn't set up to estimate skill.

 

[Bughouse]

Solutions that should make people happier about decay, though I do not know their technical feasibility in PS's infrastructure

1. Decay begins only after like the 3rd day of inactivity. Or the 5th. Or whatever. Just longer than it currently is. It keeps people from parking an alt on the top forever, because decay will eventually kick in, without harming people who just don't like laddering every day (hint people with lives)
2. Actual decay never happens, but if you haven't played x number of games in y number of days your account will not be listed on the leaderboard, so there's no bragging rights. Your ELO will only be seen by !rank, which impresses no one and in fact shows that you haven't played in a while, which you can't tell from someone who has just slowly decayed.
3. The solution Stratos/dice have already said

In other news, I want to address Antar's point that we need decay. I'm aware that PS does not use vanilla ELO. Frankly, not even chess does.

Why is rating inflation necessarily a bad thing? Chess, the OG when it comes to ELO has "suffered" from rating inflation for years partly because of rating floors, partly because of glicko bonus points, partly because there are just more people playing so the top 100 people are actually even farther from the mean person so of course they're higher rated... There's many reasons, but in any case, it doesn't make it any worse of a game. The only true negative to rating inflation is that it makes it harder to compare the quality of players across time periods. But competitive players don't really put much stock into ELO anyway. Does anyone care that the top ladder score in 2014 is lower than it is today? I don't think so.

Secondly, even if inflation truly is such a problem (which I don't agree with in any case), just deal with it by doing more frequent ladder resets (ie every 3 months at tier shifts), which was something that has been talked about plenty of times already iirc but I don't think has happened. If it is already doing this and rating inflation is somehow happening in a time as short as 3 months, then ignore this last point.

 

[Stratos]

I'm not a mathematician, though I read up on Antar's posts on the ranking systems and stuff and that's where I'm getting my arguments from. If anything I say is wrong, please correct me.

Rating inflation is actually a technical problem with Elo, and Zarel's implementation specifically (one that I think is well resolved by having actual decay).

This is a bit technical, but here goes:

• Elo as originally defined is "conservative" in that the net number of points is preserved--one player gains only as much as the other loses. This means that the average rating is always the starting rating (in our case 1000). However, even with "vanilla Elo," rating inflation is still a problem. First, new accounts mean there's consistently new points "entering the system." And the players (alts) that leave tend to have below-average ratings. So if you look at the average rating among ACTIVE players, even with vanilla Elo, rating inflation will still happen.
• PS does not use vanilla Elo. Specifically, IIRC, if your rating is above a certain point, you gain/lose fewer points than if your rating is below a certain point. This should in theory slow down rating inflation (higher-rated players will have difficulty climbing higher) but it also means that they're less likely to lose points.
• PS also, significantly, does not allow a player's Elo to drop below 1000 (this was a wise decision done to encourage players to NOT alt-reset. But this means that ladder inflation is an even more significant problem.
• Decay is the MOST STRAIGHTFORWARD WAY to deal with this problem. And it NEEDS to be an actual decay of rating, NOT just a display-thing like it was for PO. Stratos, I'm not saying your concern isn't valid, but the solution is adjust the decay algorithm not to ditch it.

My personal feeling is that Elo (ZElo especially) is likely a poor skill estimator (I'm going to put it on my to-do list to numerically analyze this), and it's why I feel we should be using GXE to rank people at the tops of the ladders. But I've been fighting that fight for a long time, and I'm no closer to winning it now than I was.

Yeah, the rating inflation is pretty clear to see. Even as just any normal competitive player, you notice it over the course of a suspect. or you can see how the more active ladders have higher peaks. My question is: is rating inflation automatically a problem? As long as players are still ranked accurately relative to one another, it shouldn't be a big deal where the peak is. Genuine question, I just don't know how inflation actually affects the ranking system.

Stratos, now let me dissect your complaint:

This is how Elo works. The higher your rating vs. your opponent, the fewer points you gain by winning. Because it's SUPPOSED to be a skill-estimator and because we DON'T want people at the top being there simply because they've played for longer. If you don't like that in a ladder (we had a LONG discussion about this a few years ago), you probably want a different ladder score system in place (I invented one called ARMS that would have suited you).

I'm 100% absolutely fine with only gaining 5-6 points per battle. If everyone I play is a +6/-34 spread, and i win roughly 34 battles for every 6 I lose, then ELO has done its job and accurately placed my skill. That's no problem to me at all. That's exactly how ELO should work. My problem is that if I have a +6/-34 spread for every battle, and I win 50 battles for every one I lose, and my ELO isn't going anywhere, then it has not accurately placed my skill.

Are you a significantly better player today than you were last week? Ratings are SUPPOSED to stabilize. They're supposed to reflect your skill, NOT your accomplishment. I wrote a big long guide about this. Again, if you want the ladder to measure accomplishment, you want a different ladder system, and one that isn't set up to estimate skill.

No, it suggests to me that my true skill level is actually higher than 1900 and my ELO needs to slightly adjust in the positive direction to compensate. I believe it was you who said that the way PS's implementation of ELO works is that you should have a roughly 75% win-rate vs people 200 points lower than you (which makes sense, as battles with 1700s players tend to give me about a +10/-30 spread). So if my win-rate vs people 200 points lower than me is above 90% because of how decay affects it, isn't there a problem with decay?

 

[Antar]

Stratos, all rating systems have to at times sacrifice accuracy in individual cases for the greater performance, so I'll need to do some analysis to determine whether ZElo has accuracy issues across the board or only for edge cases like yours.

I will say, though, that even the issue you describe seems minor compared to the benefits we get from Zarel's mods to Elo (fewer alt resets, no "parking" of alts). Ratings being off by a few hundred points would only matter if we used them for, saw, weighted stats or suspect test reqs.

 

[Zarel]

my score will go down if i lose. If I win two hundred games in a row, why shouldn't my score be in the exosphere? Every game I play I'll be looking at a +1/-39 spread and if i can keep my rating going up despite that by maintaining at least a 40:1 W/L ratio versus players near the top of the ladder, I think I've earned it. And if i try to "park" because I don't think I'll be able to maintain my ladder score, then it will start decaying and I'll be forced to play and risk losing 39 points. Your score doesn't "just keep going up and up" unless you do not lose; that's the point of ELO, to eventually find a rough equilibrium point where your wins and losses multiplied by their weights will balance out and keep you looping around in roughly the same skill level relative area. The only thing we need ladder decay for is to force people to keep playing so that ELO can do its job.

If you won 99% of your games last week, and you won 99% of your games this week, it sounds like nothing changed, so your Elo shouldn't be any higher, right?

This is a serious concern and means Zarel needs to take another look at the matchmaking algorithm.

I'm not really surprised. We stop searching past a rating difference of 300, which I didn't expect to be a serious issue at the time.

We now search to a rating difference of 600. It takes around 2 minutes to grow to 300 and around 7 minutes to grow to 600. I'm not sure if I really want to support pushing it past 600.

Yeah, the rating inflation is pretty clear to see. Even as just any normal competitive player, you notice it over the course of a suspect. or you can see how the more active ladders have higher peaks. My question is: is rating inflation automatically a problem? As long as players are still ranked accurately relative to one another, it shouldn't be a big deal where the peak is. Genuine question, I just don't know how inflation actually affects the ranking system.

Ideally, players should also be ranked accurately relative to other ladders. So you'd see "2000" and think "wow, that guy is good" instead of having to figure out which ladder it is first before you know whether or not it's good.

I'm 100% absolutely fine with only gaining 5-6 points per battle. If everyone I play is a +6/-34 spread, and i win roughly 34 battles for every 6 I lose, then ELO has done its job and accurately placed my skill. That's no problem to me at all. That's exactly how ELO should work. My problem is that if I have a +6/-34 spread for every battle, and I win 50 battles for every one I lose, and my ELO isn't going anywhere, then it has not accurately placed my skill.

(It's spelled "Elo". It doesn't stand for anything; it's named after a guy named Elo.)

That's true. But you said you're gaining 5-6 points per battle and losing 5 points a day. So as long as you're playing more than one battle a day, it should be fine, right? I fixed the matchmaking thing so it shouldn't take 40 minutes anymore.

 

[Zarel]

Doubt much could be done about it, though, getting a 'game' vs someone in 1100 just feels horrible for both parties, and if there's no one that's on the ladder in that range and you have to be matched up vs someone like that if you ever want a game, well it's going to feel bad for someone either way.

marilli, you deleted this, but I think this is a good point.

At what point should we say "I'd rather wait for someone closer in elo to me" vs "I'd rather play someone much much lower rated than me"? Do different people have different opinions?

Old algorithm would match you:

instantly: anyone within 50 Elo
after 1 minute: anyone within 180 Elo
after 2 minutes: anyone within 300 Elo

New algorithm does:

instantly: anyone within 50 Elo
after 1 minute: anyone within 180 Elo
after 2 minutes: anyone within 300 Elo
after 3 minutes: anyone within 360 Elo
after 4 minutes: anyone within 420 Elo
after 5 minutes: anyone within 480 Elo
after 6 minutes: anyone within 540 Elo
after 7 minutes: anyone within 600 Elo

I can tweak these numbers, though. Or maybe make a specific "Yes or no?" popup for whether or not they want to search past 300 Elo?

 

[AM]

marilli, you deleted this, but I think this is a good point.

At what point should we say "I'd rather wait for someone closer in elo to me" vs "I'd rather play someone much much lower rated than me"? Do different people have different opinions?

Old algorithm would match you:

instantly: anyone within 50 Elo
after 1 minute: anyone within 180 Elo
after 2 minutes: anyone within 300 Elo

New algorithm does:

instantly: anyone within 50 Elo
after 1 minute: anyone within 180 Elo
after 2 minutes: anyone within 300 Elo
after 3 minutes: anyone within 360 Elo
after 4 minutes: anyone within 420 Elo
after 5 minutes: anyone within 480 Elo
after 6 minutes: anyone within 540 Elo
after 7 minutes: anyone within 600 Elo

I can tweak these numbers, though. Or maybe make a specific "Yes or no?" popup for whether or not they want to search past 300 Elo?

I'd rather wait on someone with a closer ELO range than me than having a large gap between ELO ranges if that's what you are referring to. - One of those different people

 

[Antar]

Zarel, no matter what the rating difference, players should never be in a situation where they NEVER get matches. Especially if we penalize players for not battling.

 

[Zarel]

Zarel, no matter what the rating difference, players should never be in a situation where they NEVER get matches. Especially if we penalize players for not battling.

40 minutes isn't "never". And as you can see above, a lot of people would rather wait 40 minutes for someone closer in rating than wait 5 minutes for someone very far away

 

[Antar]

And as you can see above, a lot of people would rather wait 40 minutes for someone closer in rating than wait 5 minutes for someone very far away

This would surprise me greatly if actually true (where "a lot of people" > 25% of players good enough to qualify for suspect tests), especially since you can only search one ladder at a time (new tabs notwithstanding).

I'd like to talk to you again about the technical details of the matchmaking algorithm, but you decide whether that's better suited for email.

 

[Zarel]

This would surprise me greatly if actually true (where "a lot of people" > 25% of players good enough to qualify for suspect tests), especially since you can only search one ladder at a time (new tabs notwithstanding).

I'd imagine the ones who would rather get a match than wait are the ones who don't have a rating high enough to need to wait.

But this definitely isn't something I can be confident on. We should poll users or something.

I'd like to talk to you again about the technical details of the matchmaking algorithm, but you decide whether that's better suited for email.

I would greatly prefer PS (the Development or Staff chatrooms are a good place). You should visit way more.

 

[blarajan]

I like the pop-up option as it's the best of both worlds. I don't think a poll is required, as there is no harm in making it so players can make that individual judgment call as they may be in a different mood at the time.

 

[Zarel]

I like the pop-up option as it's the best of both worlds. I don't think a poll is required, as there is no harm in making it so players can make that individual judgment call as they may be in a different mood at the time.

It's a lot of work, though. I'd only be willing to do it if a poll was very close to 50-50.

 

[marilli]

If giving people both options is a lot of work, would it be any easier to have a popup that says "There are no opponents within your range, try again later?" and give them the option to either quit searching on the ladder or just keep searching beyond 300 elo differentials. Would this also be a lot of work?

Sorry no programmer

edit: if it's not gonna happen I'll probably content with just quitting search manually after whatever time, might still be an ok feature so just asking

 

[blarajan]

It's a lot of work, though. I'd only be willing to do it if a poll was very close to 50-50.

May I ask why? If it was unclear, I was envisioning the following, for all ladder searches, simply based off this statement from one of your posts: "Or maybe make a specific "Yes or no?" popup for whether or not they want to search past 300 Elo?"

instantly: anyone within 50 Elo
after 1 minute: anyone within 180 Elo
after 2 minutes: anyone within 300 Elo
[popup asking if they want to keep searching]
after 3 minutes: anyone within 360 Elo
after 4 minutes: anyone within 420 Elo
after 5 minutes: anyone within 480 Elo
after 6 minutes: anyone within 540 Elo
after 7 minutes: anyone within 600 Elo

Isn't that exactly what you suggested? My post was simply saying that having the popup is better than making an ideological decision and deciding between users preferring battling a worse opponent and battling at all.

Unless you're saying, you made the recommendation, but would rather not because it's a lot of work, which I can understand. If that is the case, which part about interjecting that in the matchmaking algorithm is particularly difficult? As a lay-person not at all associated with PS! programming, it seems like it would be a relatively easy prompt to code but I am likely very wrong considering your apprehension.

 

[Stratos]

man i didnt ladder for one day because i was sick and i went from 1909 to 1893 :( played a game today and went back up to 1898

 

[Zarel]

Unless you're saying, you made the recommendation, but would rather not because it's a lot of work, which I can understand.

Yes, that's the case.

If that is the case, which part about interjecting that in the matchmaking algorithm is particularly difficult? As a lay-person not at all associated with PS! programming, it seems like it would be a relatively easy prompt to code but I am likely very wrong considering your apprehension.

It means you need to interject client code into what was previously a pure function. i.e. the function goes from "it does one thing and one thing well" to "it does lots of things, and we're not entirely sure what"

Which would already be really annoying and makes separation of concerns complicated in a singleplayer game, but PS is an online game.

So now you go from a pure function to sending a signal from the server to the client, writing client code to understand that signal, writing client code to show a popup, writing client code to deal with clicking "Yes" on the popup by sending a signal back to the server, writing server code to understand that signal, and then writing server code to flag that case and having the matchmaking code handle it differently.

Not impossible, but hard enough that most of our coders will be like "omg no not client code" and hide, and I'm pretty sure I'll be the only one who even knows how to do it.

edit: actually Slayer95 can probably do it if you ask nicely