AlesCieply wrote:
Bill Spight wrote:
My preference would have been to use a different bot, such as Elf, to estimate the chunks by Carlo. But you're the one who has made the most thorough analysis, and you used Leela 11.
Just give me a bit more time, the results of the analysis with Leela Zero (ELF weights) are already in the pipeline ...
Great!
Quote:
Bill Spight wrote:
AlesCieply wrote:
In fact, in the Italian appeal they say that they got a better agreement between Carlo's moves and those suggested by Leela (top 3 choices) when they were running Leela on lower numbers of nodes. This indicates that the winrate estimates found at high number of playouts are not indicative for Carlo's moves.
I don't follow you. The assumption I a making is that Leela 11, although flawed, is able compare moves reasonably well, especially with many playouts. My preference would be to use estimates at the same depth of the tree, because of potential horizon effects, but I'll use the delta if necessary.
The moves and winrates found at 300k+ nodes might be different from those seen by Carlo when he made his move. Concerning the winrates, even at 300k+ nodes it is not rare that the winrate evaluation is off by more than 1%, though they are still determined more precisely than the top suggestions where there can be several options with a similar winrates. Carlo could have played Leela top suggestion even after some additional pondering and the played move winrate could easily drop by 2-3% when evaluated at 300k+ nodes.
I gather that when you say, "This indicates that the winrate estimates found at high number of playouts are not indicative for Carlo's moves," you mean that the winrates for Carlo's moves
that he would have seen if he were cheating are not the winrates for those moves after running Leela for at least 300k playouts. (Although after almost 4 min., quien sabe?
)
True, but here I am using Leela to estimate the degree of error. If Carlo picked Leela's top choice after running it for about one minute, and the winrate dropped enough with more playouts, so that its estimated error is more than 2%, it's still considered an error.
Quote:
Bill Spight wrote:
AlesCieply wrote:
If I was to make an assumption on why Carlo used a lot of time on some moves I would say that he was more likely to spend more time when Leela's suggestions did differ from his own instincts and in positions when he suspected that Leela might be wrong. As a regular user of Leela he must know for sure where Leela's weaknesses are so he should spend more time when such situation occurs.
In that case, we should expect that some of the deltas actually indicate that Carlo's choice is better than Leela 11's.
I think so too. The question is how often it happens. Can we see it in a limited number of positions taken from one game. And can Leela realize the played move is better as soon as it is played.
A few times in this game Leela considered Carlo's play to be better than its top choice, but not by much.
Quote:
I think the long time taken on
(your position 3) is telling. I am a mere 1d player but I believe even much stronger amateur players would just play as Carlo did (making life in a corner) but without thinking on it too much (I would not think at all on it). What caused Carlo to ponder that long. Was it not Leela persisting that there was a better move. This is a clean example in accordance with my (and Bojanic's) hypothesis that he would think longer in positions where he suspects Leela to misjudge L&D. Well, he could have just went to a toilet, accept a brief telephone call or whatever else ...
OC, we do not know why a player might take a long time on a move, but my working hypothesis is that his or her play is generally worse after taking a long time, because the position is subjectively difficult for that player. And, indeed, in this game when Carlo took 49 sec. or longer to make a play, his error rate tripled, according to Leela in your earlier analysis. (I did not check that again, with your latest analysis.)
Now, I am willing to entertain Bojanic's idea that Carlo simply ignored Leela for
, because Leela is relatively weak at L&D. But if Carlo was watching Leela's calculations, he would have noticed two things that remained fairly constant throughout. First, Leela understands that White to play in the top left corner can hold Black to one eye. Second, despite that fact, Leela prefers the eye stealing tesuji for Black in that corner (after playing kikashi in the bottom left corner or not). Leela may be looking at a possible seki in the corner, or maybe at a sacrifice. Also, the invasion at
aims at the eye stealing tesuji, especially since it would also be a two space extension from
. (Carlo was lucky that
prevented that extension cum eye stealing tesuji, defending his group on a small scale instead of attacking
on a large scale.) Surely the eye stealing tesuji was on Carlo's mind. He did not need Leela to suggest it to him.
If Carlo was using Leela to cheat at
, why didn't he follow its suggestion? It is possible that Leela preferred Carlo's play at the time he made it, but just barely. Every analysis I have seen of this game regards it as an error, and, except for this last one, a significant error. Bojanic's Leela rates it as losing 7%.
Leela's eye stealing tesuji is more than a suggestion.
Login
Register
FAQ
Search
Uberdude
this move is not played much nowadays as alpha-go joseki prefers a different one
most strong player would just play c2, looks like a bot-play
good move but easy to find, most dan players would play there, not suspicous to me
important for a shape, not sure whether 4d player would see it
strong line of play
exquisite timing, not many strong amateur players would choose the right time to play this sequence
, 
plays the eye stealing tesuji in the top left corner. 
, with 293561 playouts, for an estimated loss of 4.1%, a rather larger number than the first estimate.
. Not only by flow of the stones, but because Black seems to be ahead, and the bottom right group is his weakest. Isn't it urgent to settle the bottom right corner?
Leela Zero is stronger than Leela 11, perhaps with less precision.
- 
are all Leela Elf's top choices. (With the 100k setting.) AlphaGo likes an early 
force White low and support 
, preventing a severe attack. 
is neither necessary nor desirable, despite being joseki. Leela 11 likes 
may provide balance to the 3-3 pts. on the right. Interesting that Elf did not prefer the AlphaGo 3-3 invasion. Also perhaps because of the dual 3-3s on the right.
. Despite being joseki, Elf thinks that it loses 14%. 
. Black chooses 
instead of "a". Then 
takes advantage of the 3-3, avoiding a severe pincer.
(7 sec.) loses 7%. It certainly looks too low. 
response. 
(49 sec.) questionable, to say the least. It thinks that it loses 11.5% winrate. It is not a play that would occur to many people, I think. Elf's top choice conveys an important lesson, IMO.
on the left side. This remains a good play for White for some time.
, indicating that it was a difficult position for him, but he picked Leela Elf's top choice. 
) for 
. Not a terribly difficult choice, in my view, versus the atari, but the future is murky. (Maybe he went to the bathroom, too.) 
(34 sec.) lost 10.8%. I think that there is a good lesson, there. (Note: We are now using Elf with the 200k setting.)
stone, then 
(1 min. 5 sec.) protects against the push and cut, but it sure looks ugly, doesn't it? Elf thinks that it loses 13.6% winrate. (What was Ben David thinking about for almost 8½ minutes on