Previously I have experimented with Brackman
team strength and simple average, coming to the conclusion that averaging yields better results.
There are still a lot of complaints about imbalanced team games, thus I have tried two other strategies.
Maximum likelihood estimator: Weighted average where the weight is higher the less uncertainty a player has. This means the algorithm will try to put a similar number of inactive/newbie and active players on each team.
Minimum standard deviation: Average biased by standard deviation. This means the algorithm will try to put a similar amount of pros and nubs on each team.
Both of these turn out to perform comparably or worse than simply looking at the average skill on each team. This only takes into account the game's outcome, whereas the game itself is what really matters. Would you think one of these two might increase game quality, even if the games might end up less balanced?
To clarify the meaning of balance: The balancer currently assures that every player has a near perfect 50% win rate when playing team games. Whether those games are simply alternating stomps or actually interesting match ups isn't considered.