White wins 8 of 10, I wonder if that suggests that komi should be a little lower. Though I suppose that it could also be an idiosyncrasy in how AlphaGo plays.

Deepmind said that 7.5 with Chinese rule is as good as it gets.

If changing it by one point, it may just flip similarly in black's favour. Deepmind said that 7.5 with Chinese rule is as good as it gets.

It's not clear to me in my head that Deepmind would really know the appropriate komi (for certain). Changing the komi could have significant impact on game winning probabilities from a given board position, and could result in an altered strategy. So all of the training that's been happening through self-play might make the most sense having 7.5 komi.

It could very well be that 7.5 komi is correct, but I would think that this should be investigated more scientifically if we really want an answer to this question.

Oh, I am not surprised that AlphaGo vs. AlphaGo games are closest to 50:50 with a komi of 7.5 under Chinese rules. I don't think that Silver would make that claim without having tried different komis.

Let's call the version of AlphaGo trained with komi of 7.5 AlphaGoX. Then I agree that it's likely that AlphaGoX vs. AlphaGoX probably has closest to 50:50 win rate using komi of 7.5 points.

But would a different type of AlphaGo that plays different moves have developed if it were trained with komi of, say, 10.5? The value network would have developed differently, probably. Let's call that hypothetical program AlphaGoY.

So an experiment where you put AlphaGoY vs. AlphaGoY might end up having games closest to 50:50 with a different komi than 7.5. Because AlphaGoY plays different types of moves than AlphaGoX... Isn't that possible?

Bill,
I think 7.5 komi may very well be correct komi - just want to point out that testing with current version of AlphaGo is not necessarily rigorous.

For example, maybe AlphaGo trained with 10.5 komi plays very aggressively as black to overcome the point difference. You come up with a different program, so it remains possible that black wins more often against itself with this aggressive strategy.

Seems unlikely to me, but I just feel a single version of AlphaGo may not be qualified to generally prove correct komi.

Some trivial observations:

• Game #2 (where B287 captured 29 white stones) had the most white stones captured by black (51), but white won the game. It also had the most prisoners total (80), resulting in the fact that over 25% of the stones played in that game were captured by the end.
• Game #10 (where W244 captured 22 black stones) mirrored #2 with an equivalent difference in magnitude between black and white prisoners but favoring white, yet black won the game.

When I was starting out, I noticed that the number of captured stones was a pretty good predictor of the winner of a pro game. The player who captured more stones usually lost. (That's captured stones, not dead stones. It may matter.) Not that I made too much of that, and was not sure at all if it was generally true, but I certainly had no fear of sacrificing stones.

Until/unless we have a better description of what is meant by AlphaGo vs AlphaGo I don't think we should draw conclusions. The same program, yes, in the sense of program vs program. But the same values in the neural nets the programs were emulating? I thought these self play games were being used to train the neural net (result in the altering of cell values) so I would expect the sets of those values to be slightly different << current "best" set of values vs set that MIGHT be better >> If that is the case, we can't speak of these games indicating "white favored at this komi" unless we know that the "current" and "trial" nets were randomly assigned colors.

I suppose it takes a little bit of the sting out of losing ko fights as well (=