Deleted earlier comment due to a bug in the code.

Here's the result of a naive brute force program that assumes a random distribution of opponents (i.e. any combo is equally likely), sorted by number of wins:

185: red/blue
269: red/red
397: yellow/blue
407: yellow/red
438: red/yellow
464: red/green
471: yellow/green
483: yellow/yellow
512: blue/yellow
528: green/green
539: green/red
561: green/blue
567: green/yellow
578: blue/red
635: blue/green
646: blue/blue

The program is here: http://pastie.org/1217024 (pipe through sort -n)

It performs 30 iterations of all 16 vs 16 matchups. Note that the player that attacks first has an advantage, so doing all 16 vs 16 balances that out (everyone is player 1 as often as he is player 2).

I signed up today to comment in this thread, so don't mock me too heavily. :)

Edit: Bumped iterations to 30 and hit points to 80,000 to try to smooth out randomness in the results.