Halite Home

Matchmaking logic


#1

A few questions about matchmaking logic, because the current process doesn't seem efficient for identifying "true" skill.

  1. When / how often is a game scheduled? I see clusters of games played within periods of a few minutes and sometimes an hour or more between games.
  2. What are the parameters to the Elo model / what does the expected distribution of "points" look like?
  3. What determines the number of players in a given game, and how did you adjust the standard Elo model for games with more than two players?

#2

When / how often is a game scheduled? I see clusters of games played within periods of a few minutes and sometimes an hour or more between games.

We pick a seed player based on the square of the uncertainty in their score. This means that you will see games played relatively rapidly when you upload a new bot.

What determines the number of players in a given game, and how did you adjust the standard Elo model for games with more than two players?

Its a random distribution from 2-6 with 2 player games happening the most often and 6 happening the least. We use the trueskill library which implement's Microsoft's modified Glicko rating system. This handles multiple player matches.

What are the parameters to the Elo model / what does the expected distribution of "points" look like?

Initial mu is at 25. Initial sigma is at 8.333.


#3

Thanks @truell20.

Does the performance of a previous bot affect the ranking of the current bot?


#4

Does the performance of a previous bot affect the ranking of the current bot?

wdym "performance." If you mean time taken, no.


#5

Hah! Sorry, I wasn't clear. I meant on the topic of ranking/matchmaking: is each bot submission completely independent of other (previous) submissions, in all respects (e.g., ranking, matchmaking)?


#6

Hah! Sorry, I wasn't clear. I meant on the topic of ranking/matchmaking: is each bot submission completely independent of other (previous) submissions, in all respects (e.g., ranking, matchmaking)?

Yes.


#7

A few questions about matchmaking logic, because the current process doesn't seem efficient for identifying "true" skill.

What problems were you having with the matchmaking?


Score for ranking