Smashboards

Cool stuff! I went with Glicko because Microsoft's documentation of TrueSkill omitted important values and formula pieces behind variables. Did you end up reverse-engineering those?

That was the case a while back, but there are a bunch of implementations that exist in the wild now. There are also papers that exist which explain it more in detail now.

Ahh, I see. As far as Glicko goes, it's fairly useful, and I implement "conservative estimate" (mu - sigma) which also allows for sigma to work as a sort of point decay on the leaderboard side. Using rankings for seeding is fairly accurate among the top set of players, but I do more "fuzzy" seeding past the top 8 or so (e.g. 9-24 are seeded equally).

Generally upsets do not happen unless someone has dramatically improved or a newcomer performs exceptionally well, so the system seems accurate enough.

Hah, I do the same thing for "official rankings" as the conservative estimate. I haven't played around with seeding based on rankings yet. If you estimate the match quality before hand and seed that way it would probably be fine. If the quality is less than 0.50 then the game will be a bad matchup.

I believe seeding once sigma < 3 seems reasonable as well. TrueSkill gets you there after about 8 games with similarly ranked people.

I've implemented a system that goes back and recalculates the entire tree when an old challonge is added or a player is merged with other players (the multiple alias problem) -- Have you had similar issues? Do you also separate between computed mu/sigma based on game?