But there is a very tricky dilemma there. Black has to decide between two options:
In the first one black has solid connection to bottom which is very tempting. So tempting that I didn't take time to consider other options long enough. Unfortunately it possibly is a losing line.
Second one is tricky because it loses a direct connection to bottom. Despite that, I argue it's correct and also winning for black. It seems so that undermining the white central group is way more important than the connection to bottom.
I distrust leela's ladder on bottom. IMO it's nonsense, however, my ignorance led to leela profiting from the nonsense and finally beating me up.
Arek, shalev, lazy: thanks for your insight. I'm astonished that despite being certain that leela played nonsense move, it is still very hard to punish such mistake. What do you think is its strength?
Something similar is true in other games, even chess. Even chess engines make errors, but you've to compute next 30 exact half moves to prove them wrong. This is not practical so you'll simply end up making another error and the game is again even. If you make an error however they'll probably spot it and punish you. This is why the human players lose.
Bill, hehe, maybe my answer isn't saying anything at all. Yes obviously it wins because it makes fewer errors. The more interesting question I guess which kind of errors it still makes? Well, hard to say. Probably it misses the sequences that require more experience. But then, compared to humans, probably it's still stronger even in these cases.
Bill, me and Arek will make a serious attempt at next tournement to prevent that. I'm sure Shalev and Daniel will also make their attempts. We'll see who wins. It's even possible that someone wins against the bot but then loses against other human players and the bot will win the tournement anyway. Ahahah. We'll see! Do you want to encourage us to play better? Then set up a prize! :)
And by the way, we can learn and improve from leela play, so it's not bad news actually. It's just that we have to reach perfect play to have any chance of winning vs leela.
Leela's code is likely in almost fixed state and that it is the network that is mostly improving from further self-play/training. Only Leela's (hex) developer can say for sure how it has improved so much in the last few months. There were large changes in the code base between leela go -> leela chess, so i assume there also largish changes to make it work with hex also.
@Carroll, gzero_bot current network was trained pure supervised with quite a large network (~128 filters, 8 layers... guessing, I can't remember) from the self play generated up until July (when we had bot competition). I attempted to add in all historical LG games on top of that, but it became significantly worse, and then I tried more self play again and it got worse also. :( Not sure what is going on, but I was thinking in starting a fresh run over xmas (ie from zero data).
gzero is mostly an alpha zero clone, but a completely different code base from leela.
Little Golem