Forums / Community / Matchmaking Feedback & Discussion

[Locked] Matchmaking Feedback Update – January 8, 2018

OP ZaedynFel

  1. 1
  2. ...
  3. 2
  4. 3
  5. 4
  6. 5
  7. 6
  8. ...
  9. 8
ZaedynFel wrote:
Monitor snip- the whole OP doesn't need quoted
I realize you've written what you've written in your MM update for this week, but it's been asked repeatedly by folks throughout this thread on the forum, and numerous other posted threads throughout Waypoint: WHEN is the next season reset/changeover going to occur? You (343) historically always create 2-month long seasons. Why has the season not yet reset for January/February? Can you please give all of us over here some feedback, or a dedicated thread devoted to this.....preferably soon? Thanks ZaedynFel.
A 10% increase in skill-based result prediction is pretty spectacular. At Tracker Network/HaloTracker.com we were able to achieve marginal improvements by tweaking parameters in various skill system implementations. Our most accurate results were from a Glicko 2 based system with some custom parameters. I think we were around 73% accuracy if I recall correctly.

ZaedynFel, can you provide any specific information about the implementation? Is it still based on a Bayesian inferencing (ie. TrueSkill) or is it perhaps more neural based? An LSTM network? Just curious :)
This post has been edited by a moderator. Please refrain from making non-constructive posts.
*Original post. Click at your own discretion.
Spoiler:
Show
Since i cant find games on my main account anymore im smurfing alot and i can tell, that i have significant higher search times until i find a game the longer i played on a single smurf. So i doubt, that this system is currently not used. I am quitting games on purpose to stay on plat to find matches quickly when someone else already quit the game ... When i reach about 1000 kills in a playlist on a a smurf, it takes me about 10 minutes to find a game, even im still plat (i have kds of 3 - 4). As soon as i invite a new account (smurf) in my lobby or create a new smurf, search times drop to a few seconds.

Dont get me wrong, smurfing is a huge problem, specially since creating new accounts is for free and take a few seconds. Reasons for me for searching on a smurf are:

i cant find games on my main account in doubles, me or my teammate need to use a smurf. (happens since the weapon tuning update)
the only playlist to find a game on a champ rank is probably hcs.
i dont want to play on a champ rank by myself, since im matching teams.
i dont want to play ffa, because i receive a loss in stats when i place 2nd or worse (it should be like in h3, upper half of the lobby gets a win in stats)

There are also other reasons to play on a smurf:

Helping your friend to rank up better.
You can play on a smurf to help each other to rank up better. We are doing this to get games quick not to boost! In result we are matching low players and getting easy wins and finding games. The problem is, that a champ and a smurf (bronz - gold) in a lobby matching plats and diamonds. Just go back to the h3 system, where it didnt help too much to play on a smurf. (A 50 and a 1 in the same lobby will still match players from 50 - 40, but they will get the higher ranked players in their lobby, thats still a huge advantage for the smurf lobby. A 49 and a 1 will search from 49 - 39 instead from 50 - 39, thats also a advantage for the smurflobby because you can avoid 50s or higher players than your actual rank)

In my opinion the best way to reduce smurfing is:

give high skilled players a reason to play on their main account by themselves by introduce party matching again!

dont reward players by playing on smurfs and boosting each others up, the higest rank in a lobby makes the search radius, everyone in the seachradius gets put into a looby, the teamparing is random, so a high level players with a smurf dont get the high level teammates they matched in the search process. A win is a win, a loss is a loss! dont give or take more CSR because of the skill in the party. That can be abused by smurfing very well too.

Make the search radius for champs to diamond if no games are available, nothing is more frustrating than seaching 30 minutes for a game. U are basically stopping high players searching.

Give a win in stats in ffa to the upper half of the lobby, not just for 1st. If im by myself and i have a reason to play good games in ffa without ruining my stats i wouldnt play on a smurf ;)

Thats just my opinion, and i think there is way more to say about stats, the ranking system, the matching making system. And btw, socialplaylists should be social as well! I guess there is also still a hidden ranked system.

And sorry for my bad english, im german ;)
A quickplay Playlist is a great idea!
I think that the playlists should all be streamlined as it is so hard getting games due to players spread out over similar game types.

SOCIAL
Quickplay
Social Objective (CTF, SH, Asslt, Ball, Breakout)
Big Team Battle
Swat
Rotational Playlist (Snipers/Infection/Grifball/Action Sack)

RANKED
Team Slayer
HCS
Free For All
Doubles

WARZONE
Heroic FF
Warzone
Rotational Playlist (Mythic FF, Assault, Turbo)

Any other Playlist should be ditched I reckon. If you want to have a Grifball game all the time, start up a Custom Game and use the Looking For Group feature and the Customs Browser! That’s what they are for.

The ranked playlists should be culled too because who cares if you are a Champ Breakout player?
For the social rotational playlists each day of the week could have it's own playlist!
Monday: Snipers
Tuesday: Infection
Wednesday: Grifball
Thursday: Action Sack
Friday-Sunday: One of the playlists above
This should be a last resort considering matchmaking times are still okay in the game.
ZaedynFel wrote:
Playlist PopularityHere’s an update on playlist popularity, using the last 7 days last of playtime. Things have shifted around a little bit.

Super Fiesta
Slayer
Heroic Warzone Firefight
Warzone
Castle Wars
Big Team Battle
Infection
HCS Fall 2017
Action Sack
SWAT
Legendary Warzone Firefight
Team Arena
Doubles
Warzone Turbo
Free-For-All
Breakout
Snipers
Mythic Warzone Firefight
Warzone Assault
Grifball
It would be nice to have this list live in the game so that you know what you're getting into (Maybe also keep it in mind for H6). Also even if Grifball, Warzone Assault and Snipers are remaining on the bottom of the list, it would be a pity to remove them.
In my Opinion it's a great example how good Halo 5's sandbox actually is, with playlists like Super Fiesta, Castle Wars, BTB and Action Sack at the top or at least at the upper half of the list.
SCDF Zero wrote:
For the social rotational playlists each day of the week could have it's own playlist!
Monday: Snipers
Tuesday: Infection
Wednesday: Grifball
Thursday: Action Sack
Friday-Sunday: One of the playlists above
This should be a last resort considering matchmaking times are still okay in the game.
That's a lot of work and a lot of potential for something to break daily. Updating playlists in-game isn't really as easy as pushing a button. There's a lot of stuff happening behind the scenes and a lot of hands in cookie jars to make it all work.

Weekly/weekend lists are one thing. Daily is a bit much.

ZaedynFel wrote:
Playlist PopularityHere’s an update on playlist popularity, using the last 7 days last of playtime. Things have shifted around a little bit.

Super Fiesta
Slayer
Heroic Warzone Firefight
Warzone
Castle Wars
Big Team Battle
Infection
HCS Fall 2017
Action Sack
SWAT
Legendary Warzone Firefight
Team Arena
Doubles
Warzone Turbo
Free-For-All
Breakout
Snipers
Mythic Warzone Firefight
Warzone Assault
Grifball
It would be nice to have this list live in the game so that you know what you're getting into (Maybe also keep it in mind for H6). Also even if Grifball, Warzone Assault and Snipers are remaining on the bottom of the list, it would be a pity to remove them.
In my Opinion it's a great example how good Halo 5's sandbox actually is, with playlists like Super Fiesta, Castle Wars, BTB and Action Sack at the top or at least at the upper half of the list.
Part of why they don't tell you this is because it becomes a self-fulfilling cycle. "Oh Grifball has a low population, I'm not going to play that." which means the population stays low, etc.
This is great and all, but this system is just going to increase wait times. I know a lot of skilled players in the BTB community, and they all just want to be able to find a match. They want to play your game, but 343 likes to cater to casuals that play their game 2-3 times a month, not everyday like these players do.

My company, Ex0 Delta Gaming, is very active in btb. Last night we searched with a team of 8 for 80 minutes without finding a game, and then i had to go to bed for work in the morning. im sure after i got off and they were searching with 7 they found a game... I personally would like to see btb ranked, but i know you have "reasons" why it shouldnt be. My whole point is, if im going to have to get a hidden rank in social playlists, why not just show that rank??

Social isnt social anymore, its lobby warrioring. Damn, i just wanna play a game of halo with my buddies, why is that such a hard task??? its not my fault you made a team game and decided teams arent best for your pockets.
ZaedynFel wrote:
Monitor snip- the whole OP doesn't need quoted
I realize you've written what you've written in your MM update for this week, but it's been asked repeatedly by folks throughout this thread on the forum, and numerous other posted threads throughout Waypoint: WHEN is the next season reset/changeover going to occur? You (343) historically always create 2-month long seasons. Why has the season not yet reset for January/February? Can you please give all of us over here some feedback, or a dedicated thread devoted to this.....preferably soon? Thanks ZaedynFel.
Yeah, we've named the season "Fall Winter" to indicate it could go on longer than usual.

This was, in part, to let us both focus on the HaloWC work we needed to finish, as well the impact of the Winter holidays.

Once things settle down, we'll start making plans around the season roll, which will likely came later than sooner.
iLoch wrote:
A 10% increase in skill-based result prediction is pretty spectacular. At Tracker Network/HaloTracker.com we were able to achieve marginal improvements by tweaking parameters in various skill system implementations. Our most accurate results were from a Glicko 2 based system with some custom parameters. I think we were around 73% accuracy if I recall correctly.

ZaedynFel, can you provide any specific information about the implementation? Is it still based on a Bayesian inferencing (ie. TrueSkill) or is it perhaps more neural based? An LSTM network? Just curious :)
We're using a version of TrueSkill updated by the original team that created it, which inludes the researcher who invented / discovered the math it uses. If you've seen the factor graph from the original TrueSkill paper, just imagine adding more variables and factors into each match's graph to account for extras like parties, kills, deaths, game mode, etc. It's a more complex model, but everything is handled simultaneously in one graph. I'll go into more detail at GDC, and the Microsoft Research team at Cambridge in the UK will release a paper with all the juicy details.

The part we are still hooking up is initializing from a TrueSkill Through Time run through all of the data back to the beginning of Halo 5. Basically, we take all of the match graphs and connect them into one ridiculously enormous and deep Bayesian factor graph (equivalent to a belief network), and then propagate the results of every match back and forth through the entire network using expectation propagation.

It's pretty impressive.
@ZaedynFel: have you considered allowing players/teams to match based on something other than standard deviations?

What I mean would probably make more sense like this:

Currently, a MMR of 0 can match anyone from -1 to +1, which is 68% of the population.
A 1 can match from 0 to 2, which is 47.5% of the population.
A 2 can match from 1 to 3, which is roughly 16%.
A 3 matches from 2 to (now it gets weird, so let's just say infinity), which is roughly 2% (a.k.a. your target goal for number of people in Onyx).

What if instead, you let them match up to the MMR that gives them a, for instance, 20% population pool?

Someone on the Diamond/Onyx border could match down into the High Platinum, but I see that happening anyway because they are teaming up with lower ranked players...
Yeah, this is the type of thinking we have been considering lately. A certain percentage could guarantee a match. It would mean different skill ranges depending on your skill, but it definitely has benefits.

I know of at least one game that does this to good effect already.
When are the ranks resetting?
When are the ranks resetting?
Your answer was already provided...
ZaedynFel wrote:
I realize you've written what you've written in your MM update for this week, but it's been asked repeatedly by folks throughout this thread on the forum, and numerous other posted threads throughout Waypoint: WHEN is the next season reset/changeover going to occur? You (343) historically always create 2-month long seasons. Why has the season not yet reset for January/February? Can you please give all of us over here some feedback, or a dedicated thread devoted to this.....preferably soon? Thanks ZaedynFel.
Yeah, we've named the season "Fall Winter" to indicate it could go on longer than usual. This was, in part, to let us both focus on the HaloWC work we needed to finish, as well the impact of the Winter holidays. Once things settle down, we'll start making plans around the season roll, which will likely came later than sooner.
...and the answer was that it hasn't been settled on yet by the 343i team.
eLantern wrote:
When are the ranks resetting?
Your answer was already provided...
ZaedynFel wrote:
I realize you've written what you've written in your MM update for this week, but it's been asked repeatedly by folks throughout this thread on the forum, and numerous other posted threads throughout Waypoint: WHEN is the next season reset/changeover going to occur? You (343) historically always create 2-month long seasons. Why has the season not yet reset for January/February? Can you please give all of us over here some feedback, or a dedicated thread devoted to this.....preferably soon? Thanks ZaedynFel.
Yeah, we've named the season "Fall Winter" to indicate it could go on longer than usual. This was, in part, to let us both focus on the HaloWC work we needed to finish, as well the impact of the Winter holidays. Once things settle down, we'll start making plans around the season roll, which will likely came later than sooner.
...and the answer was that it hasn't been settled on yet by the 343i team.
K
ZaedynFel wrote:
Yeah, this is the type of thinking we have been considering lately. A certain percentage could guarantee a match. It would mean different skill ranges depending on your skill, but it definitely has benefits.

I know of at least one game that does this to good effect already.
Even though I "proposed" this idea, I still have reservations about it. I haven't sat down and done a bunch of fancy math yet, but I'd imagine one of the major drawbacks is a 0 could only match to roughly +/- 0.3 (Gold 4 to Plat 2). I suppose that would provide a fairly close match, which I guess isn't a bad thing for average players, but you might have to curve the population percentage a little at the middle.

Higher level players would be at least to find a game, which is the most common complaint I see from them. I suppose then the next complaint would be that the matches arent close or "fair", but that doesn't happen when they smurf, so I suppose it's pretty close to a net zero sum...

The other thing you'd likely have to account for (maybe only marginally) is skill spread for fireteams. I know there's the argument that the team is an average of its parts, but I don't know that I agree with that (I don't have any actual data to confirm this theory, just observed play experience that mostly contradicts this idea). Maybe the impact of this is lessened by the need to team up with lower leveled players to find a match? Like I said, I don't have any actual data to compare with...

Which game is that?
ZaedynFel wrote:
Skill SystemAs some of you saw, I’ll be speaking about the new system we are in the process of integrating at this year’s GDC. We now have just enough of the plumbing built to see results coming back from matches. The accuracy is significantly higher than what we have today despite the fact there are a few key pieces we still haven’t hooked up. That accuracy will only grow as we finish.
Have you ever applied this to evaluate past Halo Championship match ups?

If so, how accurate are the predictions? It would be a fun exercise, as well raising a lot of interesting possibilities, issues and concerns.
ZaedynFel wrote:
iLoch wrote:
A 10% increase in skill-based result prediction is pretty spectacular. At Tracker Network/HaloTracker.com we were able to achieve marginal improvements by tweaking parameters in various skill system implementations. Our most accurate results were from a Glicko 2 based system with some custom parameters. I think we were around 73% accuracy if I recall correctly.

ZaedynFel, can you provide any specific information about the implementation? Is it still based on a Bayesian inferencing (ie. TrueSkill) or is it perhaps more neural based? An LSTM network? Just curious :)
We're using a version of TrueSkill updated by the original team that created it, which inludes the researcher who invented / discovered the math it uses. If you've seen the factor graph from the original TrueSkill paper, just imagine adding more variables and factors into each match's graph to account for extras like parties, kills, deaths, game mode, etc. It's a more complex model, but everything is handled simultaneously in one graph. I'll go into more detail at GDC, and the Microsoft Research team at Cambridge in the UK will release a paper with all the juicy details.

The part we are still hooking up is initializing from a TrueSkill Through Time run through all of the data back to the beginning of Halo 5. Basically, we take all of the match graphs and connect them into one ridiculously enormous and deep Bayesian factor graph (equivalent to a belief network), and then propagate the results of every match back and forth through the entire network using expectation propagation.

It's pretty impressive.
Awesome. Those variables are pretty critical to accuracy. It's impressive you all have been able to include them all in the TrueSkill factor graph. TrueSkill through time sounds wild, looking forward to seeing it in action. Thanks for the reply!
"classic halo 5"
Meaning the standard style of play - magnum/AR. Current social playlists are things like infection, griffball and fiesta. This new playlist would be the only one that offers a standard style of play the way the game was designed, that was the point I was trying to make.
Apoll0 wrote:
ZaedynFel wrote:
Happy New Year!It’s good to be back! I have a few fun updates.
PlaylistsWe have a couple of playlist changes coming this week. First, as we announced last Friday, the HCS list will be updated to reflect the official HaloWC maps and modes. Head over to http://halo.gg for more info.

In addition, we’ll be adding a simple playlist called Quickplay at the top of our Social lists. The purpose of this list is largely so incoming players having an obvious starting point. The biggest feedback we get from new players is many if not the majority haven’t played any Halo in the past and have no idea what those playlist names mean. This should make it super clear where to start. The list will feature Slayer on Plaza (not Rig), Truth, and Coliseum. We’re keeping the maps to a minimum so folks aren’t overwhelmed, and using Slayer since it’s a straightforward game mode for both new and solo players. The playlist will also have a maximum party size of 2. If you’re looking for some light Slayer gameplay with the chance to mentor incoming folks, hop in!
Why no Rig slayer? That's one of the most fun. :'(
<p>&lt;p&gt;&lt;/p&gt;</p>
ZaedynFel wrote:
FYI folks, I put "Rig" when I meant "Plaza".

So Plaza, Truth, and Coliseum.

This should result in a few less noobs falling off maps.
Less noobs falling off maps haha. sound reasoning for an introductory playlist.
I sorts suspected that had something to do with it but, man, Rig is pretty damn fun. Oh well, thanks for replying, Zayden :)
ZaedynFel wrote:
Yeah, this is the type of thinking we have been considering lately. A certain percentage could guarantee a match. It would mean different skill ranges depending on your skill, but it definitely has benefits.

I know of at least one game that does this to good effect already.
Even though I "proposed" this idea, I still have reservations about it. I haven't sat down and done a bunch of fancy math yet, but I'd imagine one of the major drawbacks is a 0 could only match to roughly +/- 0.3 (Gold 4 to Plat 2). I suppose that would provide a fairly close match, which I guess isn't a bad thing for average players, but you might have to curve the population percentage a little at the middle.

Higher level players would be at least to find a game, which is the most common complaint I see from them. I suppose then the next complaint would be that the matches arent close or "fair", but that doesn't happen when they smurf, so I suppose it's pretty close to a net zero sum...

The other thing you'd likely have to account for (maybe only marginally) is skill spread for fireteams. I know there's the argument that the team is an average of its parts, but I don't know that I agree with that (I don't have any actual data to confirm this theory, just observed play experience that mostly contradicts this idea). Maybe the impact of this is lessened by the need to team up with lower leveled players to find a match? Like I said, I don't have any actual data to compare with...

Which game is that?
BTW, really like this general idea, radar.

It would probably be quite easy to circumvent that problem you mention. Just make the acceptable player-to-player skill gap the wider of two ranges: your idea and the current implementation. Then search times are strictly equal to or faster than today, they just get faster for people in the tails of the MMR distribution, though the MM quality could decrease out there, potentialy.

I think the biggest challenge might be how the wider allowable MMR gaps would influence team balancing after players are picked. It might not be trivial to update the TEAM allowable MMR gap if you tried to incorporate your suggestion. But I haven't thought the math all the way through, I'll admit.
  1. 1
  2. ...
  3. 2
  4. 3
  5. 4
  6. 5
  7. 6
  8. ...
  9. 8