40 replies to this topic

[Wintersdark]

Gigliowanananacom, on 21 April 2016 - 01:01 PM, said:

I didn't want to downsize the game and look up every player 1 by 1 though........ sounds like a nightmare.

Screenshot, then compare offline. For those with shown tiers anyways.

[MechWarrior319348]

Wintersdark, on 21 April 2016 - 01:03 PM, said:

It builds one team, then it builds the second. It starts building the first team before it has 24 players - it just picks appropriate players vis a vis PSR and weight class.

So you're saying that 12 players are made into a team.

Then it uses that team as a model to build the enemy team?

So its trying to "clone" the first team?

hmmmmmm interesting

Edited by Gigliowanananacom, 21 April 2016 - 01:10 PM.

[MechWarrior319348]

So I go into a match, and I end up being matched with a bunch of friends that really "click"

The matchmaker used my team as the first team. Which the Matchmaker will use as a template for the second team.

Is it possible, that as each match ends, the same people from the first "original" team will keep getting grouped into the same "original" team.

Meanwhile, the second "clone" team will always be randomly generated as an attempt to match the "original" team

Would that mean that the clone team is always being forced to draw the cards?

Is it possible that the "clone" team is "that team" that we all come across? you know.... the occasionally ****** one?

So its luck of the draw whether or not you get put into the "original" team. And once you're in, you have a good chance of keeping in that trend of a team.

Edited by Gigliowanananacom, 21 April 2016 - 01:22 PM.

[Mawai]

The biggest issue with the analysis is that you have no reliable data on the tiers of the people you were playing with. They may choose to respond correctly, incorrectly or not respond at all. In addition, if one person says "Tier 2" ... that may influence how others respond as well. So, unfortunately, your 24 hour marathon proves nothing.

Karl Berg's thread goes into a lot of the details ... including some of the checks and balances that were implemented to get things as even as possible given that the contents of the queues in terms of PSR and weight class vary significantly both on short and long term time scales. Karl was a rarity at PGI ... a communicative employee that did an amazing job of satisfying the curiosity of the community and explaining how things worked (he moved on to Amazon about a year ago if I recall). He also seemed like a very heads up developer but we didn't get to look at the code so who can tell for sure

Anyway, you can look at the Steam population cycles to get an idea of the long term queue variations ... the lower population times are when the matchmaker is more likely to expand the search parameters.

Also, the matchmaker uses the player that has been waiting the longest in the queue as the seed PSR value ... PSR range starts based on their value and expands from there as the match waits to be filled. I don't know whether the matchmaker runs a final optimization pass for solo queue matches once it has 24 players ... however, keep in mind that it must balance weight classes first so there are more constraints that may prohibit the ideal skill distribution.

You can't necessarily take 4 tier1, 10 tier2, 10 tier3 (assuming the PSR range went that wide) when each of those players is in different mechs.



P.S. I would say that the tier level for MOST players is not available ... most players do NOT post or check these forums and many of the players on the forums don't display their tier.

Edited by Mawai, 21 April 2016 - 01:36 PM.

[Aleksandr Sergeyevich Kerensky]

Actually, every player has a hidden rating number. And MM tries to get 2 teams of each rating paired up against each other within a devation of like 200 rating. After a set duration of time elapsed, if MM cannot find and pair enough players within the current teir, it expands it search +/- a teir but it still tries its best to keep the teams fairly close within that hidden rating number.

Russ even explained the system while he viewed and narrated how it worked in real time with actual people queing in the servers during a town hall a couple months back.

For the most part, it works pretty well, the only reason why it seems like it doesnt is, the nature of how 12x12 works in mwo. As mechs from 1 team start to get removed from the battlefield, the other teams advantage starts to snow ball. The firepower that team can bring to bear is compounding against the other teams lowered ability to share armor.

[MechWarrior319348]

Hmm, well its pretty interesting. Kind of makes my head hurt........ Sounds like a damn shame they lost that programmer though.

So what is the solution if you find yourself in one of those "snowball" teams. Wait 5 minutes after your last match has ended, then try to get thrown into a new pool of people?

Edited by Gigliowanananacom, 21 April 2016 - 01:57 PM.

[Wintersdark]

Gigliowanananacom, on 21 April 2016 - 01:09 PM, said:

So you're saying that 12 players are made into a team.

Then it uses that team as a model to build the enemy team?

So its trying to "clone" the first team?

hmmmmmm interesting

Well, sort of.

The actual algorithm:

1) Take the oldest player in the queue. He is the first player in this match, and the match is seeded at his PSR rating (specific number, not tier).
2) Find 11 more players, as close as possible to that rating and fitting 3/3/3/3 limitations. If matchmaking time > 1m, then allow +/- 1 tier (* There's uncertainty on how this works, more on this below). If matchmaking time > 2m, then allow +/- 2 tier. Also, the MM slips to 4/4/4/4 and 5/5/5/5, presumably at the same intervals.
3) Take the average of these players ratings to determine team average
4) Build second team seeded at that average. Note that the restriction slipping in step 2 and here are of course from the start of matchmaking.

Using this, under normal circumstances the two teams are within ~100 rating of each other, down to ~50 at peak times.

Some take-aways:

While it's easy to say that "oh, I'm high teir, so the MM slots in low-ranked players to balance me out" that is not the case. Players are added as close as possible to the target rating, so this WILL tend to bounce up and down as it stretches the restrictions both upwards and downwards simultaneously. As well, it's just as possible that it'll JUST add higher or lower ranked players, if that's what's available. It's not "trying to balance" anything, it's just grabbing players as close together in ranking as is physically possible.

At the extreme ends of the scale, you'll tend to stay in the queue longer, because you're appropriate for inclusion in less matches, and the vast majority of players are mid-ranked. So, if you're in the queue for a long time, you'll inevitably be the oldest person in the queue, and a match will be seeded at your ranking. If you're at the top of T1, this means that virtually everyone added is likely to be lower ranked than you, then the opposing team will be seeded at the average of your team, leading to you very likely being quite highly ranked vs. the bulk of the other players in the teams.

Of course, things like the above are only relevant when queue pops are lower, otherwise there ought to be enough people around your rank to fill a match. The basic truth of things is, the matchmaker can only do as well as it's inputs allow: it can't manufacture players. No amount of tweaking the algorithm is going to change that. The only options are to wait longer to build a match (and delay the amount/rate of restriction relaxation); but while that's acceptable for some, it's not for others.

Ultimately, the matchmaker is pretty simple. It works very well, when there are enough players. When there's not (Oceanic servers anytime, for example) then you're getting more and more towards "it's just a random set of players to have people play a match rather than stare at the screen"

---

* I suspect, but do not know, this is fully theory post change from Elo -> PSR, but I'm reasonably certain that the "+/1 one tier" thing is +/- a rating equal to the size of a tier. For example, if:
Tier PSR
t5: 0-200
t4: 201-400
t3: 401-600
t2: 601-800
t1: 801-1000

then one could assume that normally, the matchmaker accepts players within +/-100 rating (100 below to 100 above, so 200 - a tiers range) of the seed player, extending to +/-200 (adding one tier, split up and down) and +/- 300 (two tiers added).

It could also add +/- 200 per minute, adding a full tier up and down each time.

This gives expected functionality when you're in the middle of a tier, but means when you're at either end, you're actually already taking in players from two different tiers from the very start of the match. This would actually be ideal behavior, as it would ensure you took in players close to the seed ranking. After all, a T2 player at 790 ranking, right close to T1, would then take from 690-890 ranking rather than from 600-800. But if that IS the case, then you'd frequently have 2 different tiers in a match that hasn't spread the restrictions at all yet (but are actually close in rating, even if not in the same tier).

It's my belief that the tier number itself is purely a user-side display, only rating matters to the MM.

But either way, fundamentally the MM still works basically the same.

Edited by Wintersdark, 21 April 2016 - 02:23 PM.

[Wintersdark]

Gigliowanananacom, on 21 April 2016 - 01:57 PM, said:

So what is the solution if you find yourself in one of those "snowball" teams. Wait 5 minutes after your last match has ended, then try to get thrown into a new pool of people?

The snowball effect he mentions is combat loss grouping. It's not really about the players at all - an early death, or two, can easily turn a match between two evenly matched teams into a 12-0 wipe.

With that said, waiting 5 minutes is actually a very reliable way to get another swaft of players. Otherwise, if you keep requeuing after a match, you'll keep playing with/against the same players (though typically with the two teams scrambled, of course) as they're all within a close ranking of you, and queuing at the same time. Outliers will tend to get pulled into other matches, so the ones who keep ending up in matches with you are going to be the ones close to you in rating.

I make it a point of doing that (waiting) when there's really obnoxious people I just don't want to play with again. I'll go tinker with some builds or what have you. Alternatively, you can swap weight classes, that alone will often help.

[Wintersdark]

And, yeah, Karl Berg was a great guy. Not only did he do great work, he spent a lot of evenings discussing the nitty gritty of the game's technical side with us, and helped connect us with PGI guys who could answer questions he couldn't.

The thread is really long, but it's totally worth browsing through (if nothing else, look for the cyan dev posts); you can learn a LOT about the game there.

Edited by Wintersdark, 21 April 2016 - 02:34 PM.

[MechWarrior319348]

OOOOOOOHHHHHH i get it.

So if i am high on that skill level, and i play when all the tier 2's and 3's are on, then am i the one who's actually causing stress on the matchmaker? While taking into account that i have been playing for a decent amount of time?

Does logging out and logging back in have any effect?

Oh you mean oldest in that specific matchmaker lineup. Does the amount of matches you play during an entire play session have anything to do with it?

Edited by Gigliowanananacom, 21 April 2016 - 03:45 PM.

[jaxjace]

OP is on point with the tier grouping, its true. I would ask him to again complete a matchmaking cycle as a two man with another t1 friend. then he can compare the group MM to the solo.

As ive said in the past though, its not the tier stacked teams that matter,
its not even the coordinated smash decks groups run,

its

*******

weapon

imbalance.

[MechWarrior319348]

Doing another cycle... hmmm

Well it was fun, probably not healthy though.

It's kind of like saying, "screw my resposibilities, im gonna party for 24 hours"

Hell of a ride but you burn out fast

If Wade Boggs can drink 64 beers on one flight; then I should be able to strap into a metal death machine for 24 hours.

Edited by Gigliowanananacom, 21 April 2016 - 04:36 PM.

[MechWarrior319348]

Thanks for the responses though, very very thorough, no emotional interference.

You didn't type your responses for nothing, I think I learned a thing or two.

Edited by Gigliowanananacom, 21 April 2016 - 04:33 PM.

[XxXAbsolutZeroXxX]

Sometimes it seems like the matchmaker is designed to keep everyone at a 50/50 win/loss ratio.

If you win too much it gives you a team of 11 crash test dummies. If you lose too much it gives you a team of 11 tier 1 pilots.

That's what makes things fustrating. Sometimes it seems as if the outcome of games are predetermined and players are just along for the ride.

It might be better if the matchmaker was designed around balancing teams as evenly as possible rather than keeping everyone 50/50.

[Aleksandr Sergeyevich Kerensky]

So i posted this in your other thread... But I am unsure if you read it, anyway... Ill repost it here so you can watch the video.

Aleksandr Sergeyevich Kerensky, on 13 April 2016 - 08:07 AM, said:

http://youtu.be/G6Vsm46noJw

2:53:xx

They start talking about teirs and MM.

(I think the following only pertains to solo que)
Tier 1s are hard capped into matching up against 1-3. They cannot be matched up against teir 4-5.

Honestly, i think solo que currently works pretty well. Not every match is perfect, but most are pretty enjoyable matches considering what psr is ment to reflect.

Russ tries to explain the psr rating system at around 2:58:xx

And continues to talk about psr/mm until about 3:10:xx (at this point he changes topic).

[Davers]

Render, on 21 April 2016 - 09:22 AM, said:

Your post is pure punk...

The time he spent is just the Sample time.
The effort he put in to it was just the amount required to get the results he posted.
Go find some other forum to poop on why don't you...

Thanks OP for the info... My experience with the game tells me you are very close if not 100% correct.

OP's post is based purely on speculation without any concrete evidence. What makes him think matches are more balanced than others? How can he claim to know everyone's tiers? We have no idea what his criteria even is.

I Zeratul I, on 21 April 2016 - 07:25 PM, said:

Sometimes it seems like the matchmaker is designed to keep everyone at a 50/50 win/loss ratio.

If you win too much it gives you a team of 11 crash test dummies. If you lose too much it gives you a team of 11 tier 1 pilots.

That's what makes things fustrating. Sometimes it seems as if the outcome of games are predetermined and players are just along for the ride.

It might be better if the matchmaker was designed around balancing teams as evenly as possible rather than keeping everyone 50/50.

But we know that the MM isn't some evil Skynet punishing players for winning by giving them bad teams. The whole "50/50" thing is based on equilibrium. When you are just as skilled as your fellow players you will win 50% of the time. If you are winning more than that it means your opposition isn't good enough, less means they are better than you. Now you can argue the merits of some of the PSR scoring (is getting the Lance Bonus actually a sign of being a better player, for example) but the base concept works.

[Narcissistic Martyr]

Lack of stats, statistical analysis impossible to perform, conclusion, irrelevant anecdotal evidence.

[MrMadguy]

This is the major reason, why I don't play MWO any more.

What I want to say, is that I don't believe in snowball effect theory. I have been playing this game since Open Beta and I'm experienced enough to distinguish noobish players, who don't know what to do, and simply idle, from skilled players, who perform organized and coordinated attack. I've experienced many matches, where team had predetermined disadvantage, such as several players being AFK, but stomped other team anyway. I've experienced many matches, where 2-3 players from enemy team make a decent comeback and simply solely stomp my whole team.

What is the problem of matchmaker? It doesn't perform final balance pass, when it have all 24 players. Seeding team usually consists of players, who has longest queue times. In most cases it's players with highest skill. But there are not enough such players in queue. So other team is always being padded with middle-skill players, like me.

You know. Recently I found a 100% accurate indication, that my team is going to lose. It isn't connected with number of LRMs - it just shows, how actively my team plays. So. If my team can't lock ANY target for more, than one second - it's 100% going to lose.

Edited by MrMadguy, 21 April 2016 - 10:59 PM.

[Prof RJ Gumby]

Wintersdark, on 21 April 2016 - 02:23 PM, said:

(...)
* I suspect, but do not know, this is fully theory post change from Elo -> PSR, but I'm reasonably certain that the "+/1 one tier" thing is +/- a rating equal to the size of a tier. For example, if:
Tier PSR
t5: 0-200
t4: 201-400
t3: 401-600
t2: 601-800
t1: 801-1000
(...)

If your info on how MM work is up to date, then I would say the MM lacks one step - final shuffling. Either on the basis of total player value (that 1-1000 numbers added) or number of tier levels, it should, after the teams were formed, to switch players between teams if that can improve balance. If, i.e. if team 1 has a total of 10 000 points, team 2 has a total of 8000 points, it takes 1 higher ranked player from team A and switches him with 1 lower ranked player using same weight class from team B, to lower that gap. Or if team 1 has 3 tier1s (rest tier2), while team B has 1 tier1, it could just pick one weight class in which there is t1 on the favoured side and t2 on unfavoured side, then switch them.

Such thing would take up like 1 second, but improve team balance a lot IMHO.

I'm totally theorycrafting now, I know

Edited by Prof RJ Gumby, 21 April 2016 - 11:33 PM.

[Ex Atlas Overlord]

Wintersdark, on 21 April 2016 - 02:23 PM, said:

1) Take the oldest player in the queue. He is the first player in this match, and the match is seeded at his PSR rating
2) Find 11 more players, as close as possible to that rating and fitting 3/3/3/3 limitations.
3) Take the average of these players ratings to determine team average
4) Build second team seeded at that average.

Sooooo:

1) Tier 1 player enters Q....Team A forms.
2) Searches for T1 players....finds three....adds them to Team A.
3) Searches for T1 or T2 players....finds two T2 and adds them to Team A.
3) Searches for T1 or T2 or Tw players....finds seven T3 and adds them to TeamA.
4) Team A is done with 3xT1, 2xT2, 7xT3.
5) Forms Team B looking for an average of T2.3
6) Searches for T2 players....finds three....adds them to Team B.
7) Searches for T1 or T2 or T3 players.....finds nine T3....adds them to Team B.

You now have a match where Team A is 3xT1, 2xT2, and 7xT3......and Team B is 3xT2, and 9xT3.

Otherwise known as a roflstomp.

Did I get that right?

Edited by Ex Atlas Overlord, 21 April 2016 - 11:56 PM.