44 replies to this topic

[Wintersdark]

focuspark, on 12 January 2014 - 10:48 AM, said:

Actually, this is possible at the far ends of the ELO spectrum. Just think about the math: Find pilots +/- 10% of your ELO score, if you're at the bottom it's only the pilots 10% above you. To make things worse ELO maths tend to create a normal distribution curve making the number of pilots near the ends of spectrum few and far between.

Indeed. And thus, the further towards the end of the spectrum you are, the longer the waits. However, this just results in longer match waits, because the MM keeps stretching the bounds as matches take longer to be filled - this is how you can occasionally get really oddball players into a match. It may start at, say, within 10%, but it doesn't stay there. Just like how it tries to find a player in a mech of the same weight class, then looks for +/- one class, then further.

Quote

While I agree the poster is liking telling stories out of school, it is possible.

Possible rarely, sure. Frequently, though? No. He's full of {Scrap}.

Quote

That said: ELO doesn't create far matches, it creates fair scoring AND since we don't see or use our actual ELO modified scores it's fairly unsatisfying.

Absolutely. The MM attempts to make fair matches, but if it can't make a fair match quickly, it'll make an unfair match and consider that in how scores get modified after.

I honestly don't think anything would be improved by seeing our scores, but I understand how it's more important for some. It would absolutely cause a lot of issues here in ForumLand... though on the other hand, it would be hilarious to see all these "We lost because my team was ****" guys see that yeah, they're not as good as they thought they were.

I'm much rather see some other opt-in ladderboard or some such... but that's another discussion.

[Void Angel]

Very good reasoning, Winter; I'd actually forgotten about that Command Chair post. The only thing I can add (unless I missed the mention somewhere) is that the matchmaker system is being rewritten for UI 2.0 and the tonnage limits - which I think will alleviate a lot of the issues with the matchmaker.

Edited by Void Angel, 13 January 2014 - 12:27 PM.

[focuspark]

Void Angel, on 13 January 2014 - 12:27 PM, said:

Very good reasoning, Winter; I'd actually forgotten about that Command Chair post. The only thing I can add (unless I missed the mention somewhere) is that the matchmaker system is being rewritten for UI 2.0 and the tonnage limits - which I think will alleviate a lot of the issues with the matchmaker.

Tonnage limits will trade one problem for another. Either those many in assaults/heavies will have a difficult time placing or those few in lights/mediums will.

What we need is an incentive for people to diversify.

[Skyfaller]

focuspark, on 13 January 2014 - 12:51 PM, said:

Tonnage limits will trade one problem for another. Either those many in assaults/heavies will have a difficult time placing or those few in lights/mediums will.

What we need is an incentive for people to diversify.

There is no incentive possible as long as mechs have double bonus from DHS in engine and the weapons provide instant-convergence pinpoint accuracy... why? Because this means the mechs that can load the most front loaded weaponry, be them PPCs or ballistics, will have a ridiculous advantage over others.

If engine sinks were 2.0 capacity, 1.0 dissipation for SHS and 1.0 capacity, 2.0 dissipation for DHS inside and outside the engine then ghost heat could be removed and mechs cease to require DHS as mandatory for any build. Removal of weapon convergence; by making torso/CT weapons not converge (fire straight ahead) and only arm weapons converge removes the majority of pop-tart & front-load weapon sniping that is endemic to this game. The game returns to a more balanced TT play where weapons hardly ever hit the same armor location at once.

Those two things working together slow down the damage output of all mechs, increases the maneuvering/team work as a defining factor in a team and returns lights and mediums to the fray as effective mechs thanks to their high mobility.

[focuspark]

Skyfaller, on 13 January 2014 - 01:12 PM, said:

There is no incentive possible as long as mechs have double bonus from DHS in engine and the weapons provide instant-convergence pinpoint accuracy... why? Because this means the mechs that can load the most front loaded weaponry, be them PPCs or ballistics, will have a ridiculous advantage over others.

If engine sinks were 2.0 capacity, 1.0 dissipation for SHS and 1.0 capacity, 2.0 dissipation for DHS inside and outside the engine then ghost heat could be removed and mechs cease to require DHS as mandatory for any build. Removal of weapon convergence; by making torso/CT weapons not converge (fire straight ahead) and only arm weapons converge removes the majority of pop-tart & front-load weapon sniping that is endemic to this game. The game returns to a more balanced TT play where weapons hardly ever hit the same armor location at once.

Those two things working together slow down the damage output of all mechs, increases the maneuvering/team work as a defining factor in a team and returns lights and mediums to the fray as effective mechs thanks to their high mobility.

What if the system detected FotM builds and have them only 50% c-bills? or unique builds and gave them 200% c-bills?

[Lostdragon]

I can tell how 90% of my matches will end in the first 90 seconds. If everyone heads toward a rendezvous we have a 50/50 shot. If anything else happens you are way more likely to lose.

The first team to score a kill is also usually going to win, especially if it is an early and unanswered kill. Once a team starts dying there is usually quite a snowball effect. Even if the team loses 4 mechs but manages to badly hurt 4 enemies those hurt mechs are going to get played more conservatively by competent players and the fresh mechs will step up to the frontline. At that point it is very very hard for the losing team to mount a comeback.

[NRP]

I just have to say this has been one of the more enjoyable threads I've read here in along time. Reasoned discussion without emotional hyperbole. Bravo people!

Edit:
Regarding encouraging mech/weight class diversity, PGI could implement more lucrative bonuses for Role Warfare. Right now, the rewards system is heavily skewed towards damage inflicted. This obviously encourages people to run mechs that can pack the most firepower and also to turn every game mode into death match. Ever tried playing Conquest in a Spider where you only cap and scout? Even if you win, your payout is measly compared to the mechs that slugged it out in a death match brawl and put up 800-1000 damage with 3-4 Savior Kills. This is why we have "Assault Warrior Online".

Edited by NRP, 13 January 2014 - 02:05 PM.

[Wintersdark]

focuspark, on 13 January 2014 - 12:51 PM, said:

Tonnage limits will trade one problem for another. Either those many in assaults/heavies will have a difficult time placing or those few in lights/mediums will.

What we need is an incentive for people to diversify.

Not so much. Premade groups will all function as "average tonnage" regardless of numbers, so it's only solo puggers who are a concern. Mediums and Heavies are all close enough to the midline that they'll be roughly a wash, leaving lights and assaults as outliers.

You will probably end up with longer matchmaking times as a solo pugger if you're in a heavy assault or light light, assuming there's not an opposite member to be matched alongside you, but the solution is the same as it is currently: As matchfinding time increases, the min/max tonnage disparity allowed increases to get a match going. It'd only be a problem if the majority of people all went light, or assault. Fortunately, this isn't normally what happens. What happens right now is heavy or light premades screw things up, there's enough variety in the Regular Solo Puggers to make decently weighted matches outside premade issues, so I don't foresee this being a problem.

Lostdragon, on 13 January 2014 - 01:46 PM, said:

I can tell how 90% of my matches will end in the first 90 seconds. If everyone heads toward a rendezvous we have a 50/50 shot. If anything else happens you are way more likely to lose.

The first team to score a kill is also usually going to win, especially if it is an early and unanswered kill. Once a team starts dying there is usually quite a snowball effect. Even if the team loses 4 mechs but manages to badly hurt 4 enemies those hurt mechs are going to get played more conservatively by competent players and the fresh mechs will step up to the frontline. At that point it is very very hard for the losing team to mount a comeback.

Not always the first kill, but IMHO definitely once you have 2-0, it's extremely likely (though not certain!) that the leading team will win.

I find, though I can tell which team will win with very, very good accuracy when I see how my team moves, and compare it with how the opposing team moves. Basically, as soon as contact is made or I get visual on the layout. If one team is together, moving with purpose, and the other is sort of milling about... it's over right then. This doesn't require the mechs to be grouped into a deathball, but just not be milling all over the place like cockroaches. If they're pushing in one direction, with focus (not necessarily moving or even focus firing, but just projecting force in a directed manner), they're going to win. Milling about just means too little of the teams firepower is utilized effectively, and too few targets are presented to the opposing team - it's essentially like battles on Terra Therma in the caldera where one team falters on one of the bridges and bottles up. Too little outgoing firepower, automatic focus-fire for the opposing team.

Result? The disorganized team is instantly down a couple mechs and falls apart. Then it's just a mop up mission.

This doesn't mean the disorganized team was bad - though there'll always be that guy who insists HE is the only good player and everyone else sucks - but just that everyone had their own plan, and those plans didn't line up, while the other team may or may not have had direction, but everyone had the same idea and so just happened to work together.

[focuspark]

Wintersdark, on 13 January 2014 - 02:14 PM, said:

Not so much. Premade groups will all function as "average tonnage" regardless of numbers, so it's only solo puggers who are a concern. Mediums and Heavies are all close enough to the midline that they'll be roughly a wash, leaving lights and assaults as outliers.

You will probably end up with longer matchmaking times as a solo pugger if you're in a heavy assault or light light, assuming there's not an opposite member to be matched alongside you, but the solution is the same as it is currently: As matchfinding time increases, the min/max tonnage disparity allowed increases to get a match going. It'd only be a problem if the majority of people all went light, or assault. Fortunately, this isn't normally what happens. What happens right now is heavy or light premades screw things up, there's enough variety in the Regular Solo Puggers to make decently weighted matches outside premade issues, so I don't foresee this being a problem.

This sounds reasonable and as a COM-TDK pilot disappointing. My guess, too, is that I'll find longer than average to very log queues because I pilot a light light.

[Wintersdark]

focuspark, on 13 January 2014 - 02:29 PM, said:

This sounds reasonable and as a COM-TDK pilot disappointing. My guess, too, is that I'll find longer than average to very log queues because I pilot a light light.

Well, that depends on how many people like piloting Atlases and Highlanders, doesn't it? I suspect it'll work out just fine.

There's a good distribution of Light and Assault pilots, in my experience.

[Little Details]

If you truly believe that our current implementation of elo is as good as it gets, or only small tweaks needed, then you should be a strong advocate of making the elo score per mech chassis (or even per variant basis) than by class so you have more than 4 elo scores.

I don't see how a single person could say a locust=jenner, treb=shd, qkd=phract, aws=hgn.

When your elo score is derived from playing fotm hgn's/victors, and then you buy an aws or atl to level and don't even have the first level of basic completed and you're still playing the same group of players, it can be...embarrassing.

Similarly, for you MWO-gods, let's pretend you don't actually run a hgn/misery/victor for a minute and instead want to run an aws-9m for giggles and a change of pace, you're going to be pretty out of place...

If PGI can calculate a number 4 times, I don't see why they can't calculate it 20/40/90 times. I can't believe it is that processing intensive that it can't read a few more rows of data as part of the MM process than just pulling from that same data set with 4 numbers per player.

Edited by LT Satisfactory, 13 January 2014 - 04:15 PM.

[NRP]

@Lt Satisfactory
I think a per-chassis ELO rating is a great idea, exactly for the reasons you point out. I imagine if it is problematic to implement it would be because PGI may have to update (and test and bug fix) code each time a new mech chassis is added. Not sure if they thought this far in advance when designing the mech/matchmaking/ELO system though. I'm not a coder either. Need to talk to focuspark about that stuff.

[Little Details]

I hope PGI looks into it, NRP.

Talking specifically about non-league play games, when you run FOTM, it's self-perpetuating for a lot of players to KEEP running those builds whether it be: a) you have to to stay competitive in that PUG/4man environment just to win or b: you have to keep your standing/ego/ggclose mentality up to praise-worthy snuff.

I know I know... "there's no bad mech, only bad pilots" but in the current game landscape, you will not be able to turn the tide near as well or make a greater contribution to the team running non-FOTM - thereby affecting the W/L outcome (ergo, ELO since you're basically saying one person has the ability to affect the outcome of the game) by NOT running the best mechs for the current buff/nerf cycle of the game.

I think a lot of players would benefit from having a little fun in other mechs knowing and playing those mechs the best they can be in those particular nerf cycles knowing it won't be the end of the world and affect their uber-ELO in the mechs that count.

Players win because it brings some diversity to the game (something that will definitely be needed once tonnage limits come out). PGI wins because it gets people spending accumulated cbills on new mechs or constantly revising loadouts of owned mechs - when cbills get spent, MC looks more attractive for heros or premium time.

Edited by LT Satisfactory, 13 January 2014 - 05:43 PM.

[Wintersdark]

LT Satisfactory, on 13 January 2014 - 04:09 PM, said:

If you truly believe that our current implementation of elo is as good as it gets, or only small tweaks needed, then you should be a strong advocate of making the elo score per mech chassis (or even per variant basis) than by class so you have more than 4 elo scores.

"As good as it gets": No. However, I've yet to see suggested changes that would actually improve it without making it easy to game.
"Only small tweaks needed": Nobody has said that. As above; it's a system that works moderately well but more importantly there's nothing currently on the table that's better - and most importantly that doesn't endanger new players.

As to Elo per mech chassis, or worse, per variant: Terrible, terrible idea. Sure, it's fine for players after they've been playing a long time, but knowing how Elo scoring is calculated and what it's calculated from, going per chassis just wouldn't work. You need a LOT of matches to render results down to how much you personally contribute to victory. Very many matches. Going per chassis (or per variant!) means you'll have a whole bunch of much more inaccurate scores.

There's a better argument to be made that we should just have one score period, but that doesn't work too well because while locust!=jenner, they play a hell of a lot more alike than Locust-Atlas.

Quote

When your elo score is derived from playing fotm hgn's/victors, and then you buy an aws or atl to level and don't even have the first level of basic completed and you're still playing the same group of players, it can be...embarrassing.

Similarly, for you MWO-gods, let's pretend you don't actually run a hgn/misery/victor for a minute and instead want to run an aws-9m for giggles and a change of pace, you're going to be pretty out of place...

If PGI can calculate a number 4 times, I don't see why they can't calculate it 20/40/90 times. I can't believe it is that processing intensive that it can't read a few more rows of data as part of the MM process than just pulling from that same data set with 4 numbers per player.

It has nothing to do with calculations or processing power, and everything to do with sample size to generate a more usable score per player.

Having scores per weight class is the "lesser of evils" like much of the Elo system and how the matchmaker works in general. If you just had one score, the massive difference in how other weight classes play would cause severe problems. Likewise, if you went per chassis, you'd have a whole bunch of totally inaccurate scores because you just weren't getting enough matches in a particular chassis. Well, unless you don't play many different chassis, in which case you may as well just have one score anyways.

Anyways, you're not going to be out of place. MWO is very unique amoungst PvP games in that player skill is vastly more important to victory than mech chassis. Vastly. In tight games where skill is relatively equal (see: competitive play) then yet the difference chassis/build makes is important. But in terms of overall matchmaking in random PUG's? Player skill is orders of magnitude more important. I'm far, far more afraid of a skilled player in an Awesome than I am in some random newbie in a tooled HGN.

Finally, while all mechs play somewhat differently, there are many general similarities within a weight class with some mechs straddling/crossing the boundaries into bordering weight classes dependant on builds. Locusts aren't Jenners, but the skills you build piloting a Jenner apply to a Locust much, much more than to an Awesome, or a Jager.

[Void Angel]

WINTERSDARK! Stop calmly making my points before I can! I swear, I leave for one drill weekend, and the forums get civil....

Or maybe I've just finally ignored enough of the troublemakers...

[NRP]

Wintersdark, on 13 January 2014 - 06:09 PM, said:

As to Elo per mech chassis, or worse, per variant: Terrible, terrible idea. Sure, it's fine for players after they've been playing a long time, but knowing how Elo scoring is calculated and what it's calculated from, going per chassis just wouldn't work. You need a LOT of matches to render results down to how much you personally contribute to victory. Very many matches. Going per chassis (or per variant!) means you'll have a whole bunch of much more inaccurate scores.

Define "a lot". Just because you need some statistically significant number of matches in a chassis to get an accurate ELO score doesn't make a per-chassis ELO score a "terrible, terrible idea". I have over 100 matches in several chassis variants alone, so when looking at the chassis group as a whole, I have a few hundred matches total. If that's not enough to give an accurate ELO score for me in that chassis, then i don't know what to say except that PGI should find some other way.

[Wintersdark]

NRP, on 14 January 2014 - 07:45 AM, said:

Define "a lot". Just because you need some statistically significant number of matches in a chassis to get an accurate ELO score doesn't make a per-chassis ELO score a "terrible, terrible idea". I have over 100 matches in several chassis variants alone, so when looking at the chassis group as a whole, I have a few hundred matches total. If that's not enough to give an accurate ELO score for me in that chassis, then i don't know what to say except that PGI should find some other way.

I'd say that's probably a pretty decent amount to get a good score.

However, you do understand that that is a LOT of matches, right? That you're largely irrelevant to the whole purpose of the Elo based matchmaking, right? Except, of course, getting you out of the kiddy pool as quickly as possible.

As I said above, the most important aspect of the Elo based matchmaking is protecting The Innocents from The Veterans. After that, it's attempting to match people against other people who contribute similarly to the victory and loss of their team.

You need to have Elo scores calculated as quickly as possible. Say it takes... 50 matches, to get a moderately useful score, and 100 matches to get a good score.

A new player is going to have moderately useful scores in a couple days playing. Even playing nought but trial mechs gets him scoring information in every weight class (assuming he bothers playing them all). 200 matches to get moderately useful scores for everything.

If it's per chassis, a player spends a lot more time in the kiddy pool - every time he climbs into a new mech. This is very bad, particularly as the player gains experience.



Now, you could argue that a player, climbing into a new chassis, would start at an Elo rating for that chassis being equal to his average rating for that weight class. This starts players where the current system just has them sitting. That would work to protect the weenies.

But!

Because Elo ratings are not public, they're not Awesomeness Scores. You don't need to have scores for each chassis, it wouldn't add anything substantial to the MM. If you're a great Atlas pilot, and you climb into an Awesome, you're still going to be a great pilot. You may add a little less to the match, but a great pilot is a great pilot. What real gain is there to offset the MASSIVE slowing of calculating just how much you contribute to a victory? There's already a large portion of randomness to matches made, so having slightly different scores wouldn't make much of a difference at all.

Finally, having different scores per class would make it a lot easier to smurf. You could, say, tank your Elo rating in Awesomes just to murder newbies. You can do that now, but you'd have to stay bad at _all_ assaults, not just Awesomes.

Edited by Wintersdark, 14 January 2014 - 01:41 PM.

[Evil Ash]

LT Satisfactory, on 13 January 2014 - 04:09 PM, said:

If you truly believe that our current implementation of elo is as good as it gets, or only small tweaks needed, then you should be a strong advocate of making the elo score per mech chassis (or even per variant basis) than by class so you have more than 4 elo scores.

I don't see how a single person could say a locust=jenner, treb=shd, qkd=phract, aws=hgn.

When your elo score is derived from playing fotm hgn's/victors, and then you buy an aws or atl to level and don't even have the first level of basic completed and you're still playing the same group of players, it can be...embarrassing.

Very good idea, one of my main problems was having a new chassis be put into higher ELO games than what I wanted for a new mech, I'll edit that one into the main post with your name on it.

[Void Angel]

NRP, on 14 January 2014 - 07:45 AM, said:

Define "a lot". Just because you need some statistically significant number of matches in a chassis to get an accurate ELO score doesn't make a per-chassis ELO score a "terrible, terrible idea". I have over 100 matches in several chassis variants alone, so when looking at the chassis group as a whole, I have a few hundred matches total. If that's not enough to give an accurate ELO score for me in that chassis, then i don't know what to say except that PGI should find some other way.

What other way would you suggest?

Rating systems like Elo (which is not an acryonym, by the way) have to balance sensitivity with accuracy. If the k-factor (a measure of the amount of potential change in a win/loss) for a given range of ratings is set too high, you get wildly fluctuating scores, while if it is set too low, the system isn't responsive enough to adjust to changing player skill levels. In a typical system, the k-factor is significantly higher for new players in order to allow the system to calibrate, then becomes lower as you gain higher rankings. Here, this is how Elo works.

Now, there's 27 chassis currently in the game, and lots more on the way - requiring veterans to drop with new players on a per-chassis rating system. Speaking personally, there is a vast difference between myself and a brand new player - this isn't ego, it's just that I know so much more about the game. I know how the maps tend to work; I know what enemy and friendly 'mechs are likely capable of; I know how the team dynamic works (or doesn't) in PuGs - I know how 'mechs in general handle with and without the variant skills. None of these things are chassis-specific, and a new player's knowledge of them will be inferior - or simply absent. In fact... oh look, Wintersdark made this point while I was typing this post. I'll move on...

The other problem with a per-chassis Elo system is complexity v. returns. Quite aside from the problem of shielding newbies from the rest of us, Elo ratings would be bouncing up and down in Elo ratings like that sing-along ball from back when. This would continue for as long as you enjoyed the game enough to try out new 'mech chassis: and I doubt that the return would be worth the complexity of using the system, or of creating it. Even sandboxing the pure newbies with their own matches doesn't solve the problem, because they're still going to have to come out into the general population sometime. Given that so many player skills are not chassis-specific, I have to agree that segregating Elo by chassis is a Bad Idea - and I don't even want to think about doing it by variant.

[Evil Ash]

Void you seem to forget that the chassis in a weight class vary intensly, when I switched from Jagers to Thunderbolts, the movement of the arms confounded me for at least 10 or so games, a lower ELO for the chassis I was on would have saved me from being absolutely destroyed those 10 first matches, and maybe I wouldn't hate the Thunderbolt as I do now. Making a good impression is essential for a chassis else it will become shunned and unused. If you don't make each mech look good when you first play it, new players have no incentive to keep the chassis or variant, and will feel generally dissatisfied with the experience.