Skill ratings

A game’s matchmaker — dot-matchmaking’s DotMatchmaker, running in a lobby server — decides who plays whom. The site keeps the ratings it decides by.

The three routes are shaped after the statistics routes, with the same two rules:

• Ownership comes from the credential, never the payload. A credential scoped to anything but an app or one of its servers is refused.
• player is the credential’s per-scope pseudonym — the player key — never a site user id.

The site rates, you report

A result carries who played on which side, how much of the match each was there for, who left, and where each side placed. It carries no rating. The server computes the new ones and hands them back.

The rule is Glicko-2 as published (Glickman 2012, with the Illinois iteration for the volatility), τ 0.5, rated per match, plus the team extension below.

Teams

Each player is rated against each other side as one opponent whose rating is that side’s mean, shifted by the gap between the player and their own side’s mean, with the RMS of that side’s deviations.

So every player on a side sees the expectation the two sides’ means predict — a strong player carried by weak team-mates is not punished for a loss everybody expected — while how far each one moves still depends on their own deviation: a newcomer moves a lot, a veteran a little.

A free-for-all is sides of one.

Participation and leavers

participation

0 to 1, and weights the player’s result. Under 0.25 the player is not rated at all and gains no game — joining for the last minute should not count as a match played.

leaver

Rated at full weight with their own side placed last — and only the leaver’s row is taken from that pass. Leaving is not free, and it does not drag the team-mates who stayed down with it.

Absence

A rating’s deviation widens with inactivity — one rating period per 14 idle days — before it is used. Somebody back after a year is uncertain again, and moves like it.

Declaring the queues

{
  "playlists": [
    { "id": "ranked5", "name": "Ranked 5v5", "game": "arena", "teams": 2, "teamSize": 5, "ranked": true, "placementGames": 10 }
  ],
  "ts": 1790000000,
  "nonce": "…"
}

{ "ok": true, "created": 1, "updated": 0 }

Up to 25 per request and 50 per owner; past the owner’s ceiling, refused[] names the ones not taken.

An unranked queue is known and never rated: its results are accepted and change nothing. That lets a matchmaker file every match the same way and leave the decision to the playlist.

Filing a match

{
  "playlist": "ranked5",
  "matchId": "3f9c2a…",
  "sides": [
    [ { "player": "k7Qx…", "participation": 1.0, "leaver": false } ],
    [ { "player": "Zp2m…", "participation": 0.6, "leaver": true } ]
  ],
  "placements": [1, 2],
  "ts": 1790000000,
  "nonce": "…"
}

{ "ok": true, "ratings": { "k7Qx…": { "rating": 1662.3, "deviation": 290.3, "volatility": 0.06, "games": 1, "last_played": 1790000000 } } }

ratings holds everybody the match rated — a player under the participation floor is absent. The keys are DotMmRating.to_dict’s, last_played in Unix seconds included, so the addon stores the answer as it comes.

placements has one entry per side; lower is better, and equal is a draw. At least two sides, at most 64 players per side and 128 in all, and a player may appear once.

Reading ratings

GET /api/integration/v1/rating/players?playlist=ranked5&players=k7Qx…,Zp2m…

{ "ok": true, "ratings": { "k7Qx…": { "rating": 1662.3, "deviation": 290.3, "volatility": 0.06, "games": 1, "last_played": 1790000000 } } }

This keeps the matchmaker’s own store fresh before it forms a match.

• As stored, not aged. last_played travels with each rating, and the matchmaker ages on read, as it does for its own store — so the widening is applied once, in one place.
• A player with no rating on the playlist is absent, which a matchmaker reads as “new”.

Read is a separate scope from write for the reason every pair is: a lobby that only queues players should not be able to file results.