MMR Matchmaking Planning Thread. Nerds welcome.

[qan]

Teams should still be balanced in pairs in every arena. Use !st / !showteams on MMRBot to see the distribution.

I added an experimental option to create tiered teams for hosts: !mtt / !maketeamstiered This will group players together by general skill level, but at the cost of some balance. Still, ratings adjustment is based on the balance of the teams, so this isn't that big an issue. Technically any two teams can play one another and the ratings change will be relatively fair. Note that new unrated players who are far above/below 3000 will especially skew this.

Something's wrong with the adaptive class loader so it's not loading the new changes for some reason, but this will be in eventually.

[Cape]

While this is being manually hosted, and before rating players is tested, public ratings and leaderboard should not exist, IMO.

[Riverside]

Bot Note: Instances when multiple games run simultaneously; Bot should place all higher rated players in the same game.

I like this idea

Dis cud add another type of competitiveness and individual improvement

[hellkite]

Bot just benched me 4x in a row. WTF!!

[PapaRaazi]

Hi qan,

1st off: nice work on MMR, ive been inactive for a long time and so far its been the best way I could get some playtime in years

2nd: are you still open to suggestions on how elo points are awarded? Or has this been decided on long ago? (havent read every post in this thread)

[Rab]

Would be good if the ladders could go next to each other, here's the code for it:

[qan]

Bot just benched me 4x in a row. WTF!!

Sorry about that. There's still code to increase "playability" based on if you were benched last game, but only if it was of the same gametype. It shouldn't happen all that often.

1st off: nice work on MMR, ive been inactive for a long time and so far its been the best way I could get some playtime in years

2nd: are you still open to suggestions on how elo points are awarded? Or has this been decided on long ago? (havent read every post in this thread)

Thanks. Glad to hear it.

We're pretty locked into just W/L MMR. Anything else requires a lot of careful planning and testing, and we don't really have the luxury of either. Still open for suggestions about uncertainty and base factor (which is K factor in Elo). In the future we might look at certainty decay if players don't play for long periods.

Would be good if the ladders could go next to each other, here's the code for it:

Thanks, Rab. Good call. All set.

[paradise!]

Yeah man just wanted to say that's awesome code you've got set up. The ELO system looks crisp!

One quick suggestion. It may sound crazy but hear me about. We could optimize it by possibly shoving it up Zeebu's ass?

[Rab]

Looks like the website needs to show more rows (all of them ideally). Would be handy to add a leftmost column for ladder position too.

[qan]

Ladder position would be easy enough. Should we be showing all rows? There's certainly some shame at being dead last with an awful rating. You can still use !r to check individual players on MMRBot.

[Rab]

Yes we should be showing all rows. It's the most convenient way to:
- see that MMR is working properly
- use MMR as input into TWDT star ratings
- see own/other's MMR at a glance

[banzi fuq]

I was under the impression that we are only in the MMR testing phase and that others provided counterpoints for even showing ratings at all at this point. I assume devs can get to the rest of the data if needed to make sure things are working. Right now showing a bunch of rows with rank numbers but without more context probably isn't the best way to improve drafts. Sure once we go official show everything including confidence levels for individuals, updated box scoring, and whatever other useful stats are being collected.

[Riverside]

I appreciate the visibility of mmr rankings, i think its a helpful tool towards incentivizing games- which could increase test data.

Maybe other #s can be displayed alongside the individual mmr ratings on the website:

• Average rating of opposing team
• K : D
• Knockouts (first/second out KOs? Total KOs? Per game?)
• K : B (%)

[qan]

Other ratings are slated to be displayed, but that's webdev work, and out of the scope of what I'll be able to do for the project.

---


There's been a slight adjustment to how uncertainty works: ratings changes will happen slightly more gradually in your first few matches. However, the curve is overall not significantly changed, so if you've played more than 10 matches, you won't see that big of a difference.

The reasoning behind this is that the initial uncertainty curve was designed with the idea that we would have qualification matches that only players with <5 games will play. When it became clear this wasn't feasible, we just threw everyone together into the same matches. This is why new/unrated players can throw off a match balance, whereas under the original plan, they'd have already been sorted to somewhere between ~2000-4000 after their quals. It's not a perfect solution, but for how we're testing at the moment, it's necessary.

A long-standing bug involving players being added to the wrong match should also be fixed. That helps alleviate the impact of a bug where players who are not in the match are being considered part of the team, throwing off ratings adjustments. (Currently, there are few sanity checks, so if a player is not in the game but was added, make sure they get removed to avoid skewing ratings.)

[Vehicle]

tl;dr: make a public schedule for mmr game times, it will boost participation.

---

I would like to humbly suggest, again, that we try to set up some kind of weekday schedule.

The efforts of beam and others have shown there is demand for weeknight dueling. There could be some debate about how and if mmr is canabalizing twbd activity. Or you might not care about that but log on to try to do mmr warbirds but it's javs that is being hosted.

I think activity, and perhaps even competition, would be boosted if we did a simple public schedule of 1 or 2 hour blocks on weeknights, perhaps ending in time for a BD or two or having a break in the middle for basing.

This would obviously require a commitment from staff, and that's really the main hurdle. But if we got a few more people, perhaps even nonstaffers, trained on the bot I don't think it would be an issue.