Computer Go and Komi?

[judicata]

Does anyone know of any recent experiments regarding what today's go-playing programs "think" (using the term loosely) komi should be?

Before you jump on me, I realize there are several reasons not to rely on computers for determining komi--or even consider any program's input--and different programs may reach different results, etc. I just think it would be an interesting experiment. To be clear: I'm not at all suggesting the computer programs should influence komi. This is just an experiment I would find interesting.

The basic idea is simple*: have two instances of the same go program play each other several times, and see what komi leads to a 50/50 result. The programs would preferably run on identical hardware, but I'm thinking of this mostly for amusement/entertainment purposes, so no need to be picky, it just makes interpreting results more difficult. For all I know, the results would be completely ambiguous or indecipherable (e.g., the black/white win rate could be 50/50 regardless if the komi is 0.5 or 10.5). Using different engines could also be interesting, but more complicated.

*Of course, the implementation of the idea could be exceedingly complex, depending on whether you want to do anything with the results other than amuse yourself. I don't.

[Suji]

Does anyone know of any recent experiments regarding what today's go-playing programs "think" (using the term loosely) komi should be?

Before you jump on me, I realize there are several reasons not to rely on computers for determining komi--or even consider any program's input--and different programs may reach different results, etc. I just think it would be an interesting experiment. To be clear: I'm not at all suggesting the computer programs should influence komi. This is just an experiment I would find interesting.

The basic idea is simple*: have two instances of the same go program play each other several times, and see what komi leads to a 50/50 result. The programs would preferably run on identical hardware, but I'm thinking of this mostly for amusement/entertainment purposes, so no need to be picky, it just makes interpreting results more difficult. For all I know, the results would be completely ambiguous or indecipherable (e.g., the black/white win rate could be 50/50 regardless if the komi is 0.5 or 10.5). Using different engines could also be interesting, but more complicated.

*Of course, the implementation of the idea could be exceedingly complex, depending on whether you want to do anything with the results other than amuse yourself. I don't.

There might be some issues with this. Let me explain. Where are you going to find two exactly evenly matched programs? Self-play has it's own issues, like having having one set of strengths and another set of weaknesses. How are you going to be sure that those strengths and weaknesses don't influence the outcome?

Nevertheless, such an experiment would be incredibly fascinating. How does komi get decided anyway?

[shimari]

See, komi always bugs me. I am quite sure that a professional player can get 7.5 points out of playing first (in AGA rules), I am equally sure that my odds of doing so are minimal at best! Komi is set by comparing the win/loss ratios of professional players, based solely on what color they are playing with. If black wins more than 50% of the time, they increase the komi. It seems to only go up, I have never heard of it going down.

[amnal]

See, komi always bugs me. I am quite sure that a professional player can get 7.5 points out of playing first (in AGA rules), I am equally sure that my odds of doing so are minimal at best!

Looking at this another way, your move quality is probably random enough that the komi doesn't have that much of an effect overall. If komi doesn't make sense to you for the reasons you suggest, it's likely that it also doesn't actually harm anything.

Komi is set by comparing the win/loss ratios of professional players, based solely on what color they are playing with. If black wins more than 50% of the time, they increase the komi. It seems to only go up, I have never heard of it going down.

This is a simplification of what has really happened. Originally, handicaps were done by having different ratios of games playing white or black, and komi is a relatively recent idea. Partly because it was controversial, it crept in at low numbers. At that point, there was really no way to know what komi would be reasonable, because current styles of professional play were very much built around the absence of it.

The process whereby komi has increased has (as far as I can tell) been mostly influenced by what seems honestly reasonable. As professional styles have adapted to komi's behaviour (and alternating colour handicaps have disappeared), it has become more clear what komi is most reasonable, so it's slowly grown as it becomes less controversial to do so. The now-popular 6.5 seems to strike quite a good balance, and I think there is anecdotal evidence of 7.5 (in Japanese rules) being slightly wrong in the other direction, so it seems like we have probably found a good value.

That said, I'm not expert, and I'd be interested to see if other people think it is likely to change any time soon.

[maproom]

For the 9x9 bot tournaments I run on KGS (the next one is this Sunday, Feb. 5th, see http://www.gokgs.com/tournInfo.jsp?id=672) I have been asked by the programmers to use komi of 7. This is to minimise the advantage to either player. They use Chinese rules.

I suppose you could try to measure the correct komi by setting a bot to take hours evaluating its winning chances on an empty 9x9 board with 6 komi, then 6.5, then 7, etc.

[xed_over]

For the 9x9 bot tournaments I run on KGS (the next one is this Sunday, Feb. 5th, see http://www.gokgs.com/tournInfo.jsp?id=672) I have been asked by the programmers to use komi of 7. This is to minimise the advantage to either player. They use Chinese rules.

Since they're using Chinese rules, shouldn't the komi be an even number (either 6 or 8) to prevent ties? Because there's an odd number of points.
http://senseis.xmp.net/?WhyChineseKomiS ... PointSteps