Hauser, on 13 May 2013 - 03:50 PM, said:
1. Step 4 isn't explained anywhere but if you think about the problem for two minutes you'll figure out that constructing both teams at the same time is the only feasible solution.
I think there are more solutions which are feasible, e.g. fetch 16 players off the pool (using a target value) and afterwards assign them to the teams such that the team average elos are similar.
Hauser, on 13 May 2013 - 03:50 PM, said:
3. You seem to be confusing a few things here. The match maker starts with a value in mind. While composing a team for this value it does not change (it just doesn't make sense to do so). What can change however is how far the match maker will stray from this value. The average elo of the teams under construction does not come into play untill they have been created to calculate the odds of winning, this chance is used to adjust players elo based on the outcome.
IMO it does make very much sense to adjust the target value according to the average elo of the players you've already assigned to a team; especially if an extraordinary elo players has been assigned. This way you ensure while composing the teams that both teams end up with a similar elo value - which is IMO the whole point of elo team MM.
Hauser, on 13 May 2013 - 03:50 PM, said:
The match maker also creates matches for different elo levels, e.g. it picks different start values (there would be no point to match making otherwise). I don't know how these values are picked. An easy way would be to use the value of the first player in the queue (this avoids having to make predictions about how many players of which elo will be online at any given time).
If there are different target values, then there are "soft tiers", i.e. high elo players would more likely be assigned to matches with other high elo players.
Still, I think there is a point to MM even if the target value is fixed; even if it's not adjusted in the way I mentioned above:
AFAIK the purpose of MM is to match two teams in a way that it's a hard game for both sides (no steamrolling). That is, to match them such that the probability/chances to win are even for both teams.
Elo is used in an algorithm to determine how probable it is you'll win against another player. It looks like PGI have extended this algorithm to compute the probability of a team winning against another team based on the elo values of the individual players. Furthermore it looks like this extension is simply to use the PvP algorithm on the elo average of the teams.
Therefore, it looks to me like MM is trying to match teams with similar average elo. If you just pull in players in the way you described, from a bracket around a target value (or with a certain distribution around this target value) you'll end up with two teams with different elo values. They'll be somewhat similar, but nowhere near
as similar as possible. This wouldn't be IMO a good way to achieve the aforementioned purpose.
I imagine you could use a priority queue to order players according to their elo's deviation from 1300, where players that wait long (having an extraordinary elo) will get prioritized eventually. This way, you'll eventually pull in a player with an extraordinary elo even if the target value is fixed. By now using the team average elo to level the two teams, you'll still end up with what your prediction says will be a "good match".
Another idea would be to adjust the target value according to the average elo of players currently in the queue. It wouldn't change much in short-term, but it wouldn't be fixed as well.
Now talk about over-thinking
Edited by Phaesphoros, 13 May 2013 - 04:25 PM.