Life In 19x19

I managed to make KataGo analysis work on my pc, by gathering a few bits from the github pages of katago and lizzie

Now KataGo has a 'score estimation' network, and who would have thought - the evaluation of an empty board with chinese rules and 7.5 komi is a 1.2 points advantage for white.

Someone might correct me, but from my understanding, Chinese rules with 7.5 komi is equivalent to Japanese/Korean with either 6.5 or 7.5 komi depending on who gets the last dame.

If KataGo is 'correct', does this mean that 6.5 komi chinese rules is actually more appropriate than 7.5?

6.5 komi chinese rules is actually more appropriate than 7.5?

The odd number (5,7,etc.) is related to the odd-numbered 361 board.

There is little difference between 5.5 and 6.5 for Chinese. I'm not sure why they switched from 5.5 to 7.5 but I guess there were some statistical reasons.

Otherwise common sense would predict 7.5 is worse: correct komi is 7 and it is better to let B win ties than W, since practical players are not close to correct/perfect play thus need slightly less komi.

Yakago wrote:Now KataGo has a 'score estimation' network, and who would have thought - the evaluation of an empty board with chinese rules and 7.5 komi is a 1.2 points advantage for white.

However - if you repeat the experiment with zero komi, it thinks Black has a 9.1 point advantage on the empty board.

bernds wrote:

Yakago wrote:Now KataGo has a 'score estimation' network, and who would have thought - the evaluation of an empty board with chinese rules and 7.5 komi is a 1.2 points advantage for white.

However - if you repeat the experiment with zero komi, it thinks Black has a 9.1 point advantage on the empty board.

But if white tries to win with zero komi, he has to try risky things, which leads to greater disadvantage if black counters well. If instead of trying to win, white tries to minimize the score difference, you may get different results. I think the average score difference would be smaller.

Yakago wrote:Chinese rules with 7.5 komi is equivalent to Japanese/Korean with either 6.5 or 7.5 komi

You are wrong, see viewtopic.php?p=246387#p246387

Have more experiments been carried out? Are there more conclusive statements on fair komi by looking at LZ's probabilities or KataGo's scoring?

Knotwilg wrote:Have more experiments been carried out? Are there more conclusive statements on fair komi by looking at LZ's probabilities or KataGo's scoring?

We need new and fair komi!!!

When I run strongest Leela on my super machine:
12944050 playouts (analysis stops every time automatically and this is detritus) = more playouts are not possible:(
Black: 43% chance to win.
White: 57% chance to win.

Or in my words to be more clear 40% vs 60% or (4/10 vs 6/10) vs (40 = 66% vs 60 = 100%).
I think your know when I mean;)

A single 40:60 statement isn't worth a lot in my understanding. The question is what happens if you lower the komi. If you lower the komi by one and end up with 70:30, the original komi was a fairer one.

I don't think it is possible to achieve 50:50 winning chance anyway. And the stronger the bots get the more skewed will be the winning chances on the empty board. The question is, what komi leads to the least skewed winning chances. It will depend on the ruleset, and I expect that button go will turn out to be the optimal solution.

Back in 1977 (if I remember the date correctly) the American Go Journal published an article that claimed, based upon statistics published by the Nihon Kiin of 1400 recent pro games at 4½ pt. komi and 1400 at 5½ pt. komi or 5 pt. komi with White winning jigo, that the correct komi is 7. Terry Benson sent me the article to review before publication, along with the data. For each set of 1400 games the komi which came closest to a 50:50 split was 6½ pts. As I recall, the percentage difference in wins for a one pt. difference in komi around a 6½ pt. komi was less than 3%, That proved to be enough of a winrate difference for the Japanese to raise komi to 6½ pts., but it took them more than 25 years to do so.

goame wrote:

Knotwilg wrote:Have more experiments been carried out? Are there more conclusive statements on fair komi by looking at LZ's probabilities or KataGo's scoring?

We need new and fair komi!!!

When I run strongest Leela on my super machine:
12944050 playouts (analysis stops every time automatically and this is detritus) = more playouts are not possible:(
Black: 43% chance to win.
White: 57% chance to win.

Or in my words to be more clear 40% vs 60% or (4/10 vs 6/10) vs (40 = 66% vs 60 = 100%).
I think your know when I mean;)

What happens with a 9½ pt. komi?

goame wrote:We need new and fair komi!!!

When I run strongest Leela on my super machine:
12944050 playouts (analysis stops every time automatically and this is detritus) = more playouts are not possible:(
Black: 43% chance to win.
White: 57% chance to win.

Or in my words to be more clear 40% vs 60% or (4/10 vs 6/10) vs (40 = 66% vs 60 = 100%).
I think your know when I mean;)

AlphaZero confirms this. In the 1000-game match between AlphaZero and AlphaGo Zero, White won 57.6% of the time, as reported by DeepMind.

(However, there must be a slight mistake in their numbers. For example, they say AlphaZero won 68.9% of its games as White, but that would mean 344.5 games out of 500. Where does the half game come from? They don't report any drawn games, and supposedly superko was used. Same thing with AlphaZero's games as Black. The decimal place in the percentages should be an even number.)

With fractional komi the strongest bots will always show skewed winrates. That unfair half point does matter at high levels.

jann wrote:With fractional komi the strongest bots will always show skewed winrates. That unfair half point does matter at high levels.

I don't think that makes a difference at all. Why would it be more likely that 7,5 results in skewed winrates than 7?
Suppose that a komi of 7 gives Black 55% and a komi of 8 gives 45%. Isn't it then to be expected that a komi of 7,5 will give 50%?

Knotwilg wrote:Why would it be more likely that 7,5 results in skewed winrates than 7?

Because top bots are within a few stones to perfect play (which gives 50% with integer komi). In any case, as bots gets stronger, any komi flaws will get amplified towards 100% or 0%.

Suppose that a komi of 7 gives Black 55%

This doesn't seem likely at today's bot levels already. Actually, hard to imagine at any level.