Balancing CTF?

[matastrb]

Assume that we already have a set of players that are divided into a number of groups:

• very good players A
• good players B
• average players C.

For each group we divide the number of players into two (almost) equal sub-groups and assign each subgroup for each team. For example players in A are divided into subgroups A1 and A2 which correspond to the red and blue teams.

If a new player enters and if his information is available then we can assign him to a team based on the group that he belongs to and based on the current balance situation inside the group. If he is in group A and A1 has 1 more player than A2 then assign him to A2, otherwise flip a coin. If he/she is completely new to the game (no info in the set of current players) then assign that player into group C similarly.

Now come the group dividing step. Assume that we have collected data from players from time to time. Each player (or IP) has a set of (averaging over time) attributes (points, captured flag, lift flags, kills, deaths, etc). The following approach is based on the belief that the groups A,B, C are well separated in high-dimension space.

Since the group labels are unknown at the beginning, we can use a simple unsupervised learning approach such as K-means to divide players into groups. The calculation can be done offline and creates a look-up table for players.

My intuition can be completely off though.