Determining the right komi in the game of go (for 19x19 board)

[fluidistic]

I've just thought of a way to get the right komi in the game of go.
Let 2 programs that plays randomly play each other millions of times, for different values of komi. A game should take less than 1 second I guess since there's basically almost no computing process for each move.
Then "make statistics" out of those games. For example, for what komi value is the higher % of jigo reached?

Is this realistically doable? Would that be accurate?


For those who don't know what the komi and jigo are:
Komi is the points the player that plays white gets at start to compensate that black plays first. The "right" komi should be the komi that leads to a draw (called jigo) if both players play a perfect game.
To give you an idea, komi is generally taken as 6.5 in KGS go server. They don't take an integer because they want to avoid draws.

 

[Greg Bernhardt]

I've just started learning to play Go so I am interested in this too! Any ideas?

 

[mfb]

Random play doesn't show the advantage of a player with non-random play.
If it would, you would not have to test different komi values, you could check the distribution of point differences and find the median. This doesn't work with proper play however, as the players can take the komi into account in their moves.

AlphaGo played many games against itself, it should be able to find some reasonable komi value for its level. Other levels will need other komi values.