Life In 19x19http://lifein19x19.com/ Video series: Learning from AlphaGohttp://lifein19x19.com/viewtopic.php?f=10&t=15319 Page 2 of 4

 Author: gennan [ Tue Jan 02, 2018 7:55 am ] Post subject: Re: Video series: Learning from AlphaGo djhbrown wrote:daal wrote:I thought the explanations were clear and made senseagree - so much so that they got me thinking and asking a question At the moment, my understanding of gennan's dia 5 (reproduced below) is: black 3 could have two meanings (i didn't mean to suggest that black should only consider the top right after any white 4).Black 1 and 3 ask white a question; white's answer at 4 also has two meanings - it either implies she thinks the outside is important, or it's a feint to busy black outside so white can take the inside... so black will reassess things and look for a move to keep her options open, making use of both black 1 and 3. Attachment:example4.php.pngI can only guess what AlphaGo and pros are thinking (from the continuations they play), but I see only one response by them after white 4 (=move 5 at E14). When only one move is possible for move 5, I see white 4 as a forcing move (or a rethorical question) rather than a question.I kind of understood that you wanted to considering other options for black's move 5 to take the outside instead of the inside (so that move 4 can be considered a question instead of a forcing move), so I tried to find a reasonable alternative for move 5. But then from your later posts it seems that perhaps you weren't even asking me a question (or perhaps you were, but I got lost somewhere). Perhaps you just wanted to confirm that move 4 is a forcing move with multiple meanings (instead of contemplating black's options for move 5)?If that is the case, then I would say: yes, move 4 has multiple meanings. It's a kind of double attack. This kind of double attack is usually called "leaning" or a "leaning attack".

 Author: gennan [ Wed Jan 03, 2018 12:19 am ] Post subject: Re: Video series: Learning from AlphaGo djhbrown wrote:.. AT says Alfie prefers F17 to white 4, despite her win% of it being much less.You are not the first to miss this, but all winrates in AT are black's winrates (as stated in the legend below the board and as I explained at the start of my video #1). So white's precentages are 46.1 for E15 and 51.1% for F17. AG prefers F17 because of that rather big difference.My conclusion in my video #1 is that E15 a rather big mistake by white. This is surprising, because E15 has been considered a joseki move for many decades. These kind of surprises are the reason I'm making these videos.

 Author: gennan [ Wed Jan 03, 2018 12:31 am ] Post subject: Re: Video series: Learning from AlphaGo djhbrown wrote:Follow-up question: what, if anything, can AT teach me about white's invasion? at the time, i thought it was on the standard invasion point, but black went after it like a ferret; in retrospect i would say it was my losing move, and i should have prepared an escape route before jumping in so deep. i had played it as a sacrifice probe, but fighting spirit got the better of me and when a long thin straggly black group running into the centre finally made eyes, black's influence was overwhelming.Attachment:invade.pngIf the game didn't go as expected after this, it does not mean that this move was the cause. If you'd like a review, I suggest you upload the game to https://gokibitz.com

 Author: djhbrown [ Wed Jan 03, 2018 1:36 am ] Post subject: Re: Video series: Learning from AlphaGo gennan wrote:all winrates in AT are black's winrates Doh! I should have guessed that a mindless moron like Alfie wouldn't know that sensible readers reasonably assume a label on a white stone says something about white gennan wrote:E15 a rather big mistake by white. This is surprising, because E15 has been considered a joseki move for many decades. These kind of surprises are the reason I'm making these videosHindsight is a wonderful thing, but now that you have shown us that Her Royal Alfness has pointed it out, E15 does look kind of clumsy, perhaps an overreaction to G16 and unnecessarily confrontational, in the psycho "cut before you think" mould, which seems to be how almost all my opps play. That i invariably succumb to such brutish tactics only encourages them, but the queen shows us that deftness beats brute force, which is kind of ironic, seeing as she uses brute force inside her head to come up with a delicate touch on the board. Maybe it's just that far-sighted brute force beats short-sighted belligerence.gennan wrote:If you'd like a review, I suggest you upload the game to https://gokibitz.com thanks for the offer; i had another look at it and saw that yes, it wasn't that move which was the beginning of the end, but an elementary tactical blunder a few moves later... and then to make matters worse, i completely forgot that my blunder meant that the thin straggly black group wasn't eyeless at all, because it had connected by cutting me where i forgot to protect. so all in all, too dumb a game to test anyone's patience to review. and it's not as if the talk by Rudd i was watching at the same time was anything worth listening to.What's the name of that disease where you forget things?

 Author: dfan [ Wed Jan 03, 2018 6:36 am ] Post subject: Re: Video series: Learning from AlphaGo I prefer the win rate to always be from the same point of view (e.g., Black), regardless of whose turn to move it is. I hate analyzing with Leela and having to spend half a second every move figuring out how much the evaluation really changed when the win rate went from 57% to 42%. Crazy Stone does this right (from my perspective).

 Author: gennan [ Thu Jan 04, 2018 12:50 am ] Post subject: Re: Video series: Learning from AlphaGo djhbrown wrote:gennan wrote:all winrates in AT are black's winrates Doh! I should have guessed that a mindless moron like Alfie wouldn't know that sensible readers reasonably assume a label on a white stone says something about white Yes, I also found this presentation choice by DeepMind a bit surprising and from comments I understand that it was counterinituitive for others as well. I was used to the Waltheri database presentation which shows the winrates for the player to move. So I had to do conversions when comparing Waltheri winrates with AT winrates for my video #2. First I thought that DeepMind made an error against the principle of least surprise, but dfan actually prefers the AT presentation. Perhaps he has a point. Initially it is a bit surprising to many people, but when one gets used to it, it may be more convenient like this.

 Author: Baywa [ Thu Jan 04, 2018 2:59 am ] Post subject: Re: Video series: Learning from AlphaGo gennan wrote:djhbrown wrote:gennan wrote:all winrates in AT are black's winrates Doh! I should have guessed that a mindless moron like Alfie wouldn't know that sensible readers reasonably assume a label on a white stone says something about white Yes, I also found this presentation choice by DeepMind a bit surprisingFrom a developer's point of view this makes a lot of sense. They chose blacks winning percentage as target variable p. Black tries to maximize p ( p->100%) and White tries to minimize it (p -> 0%). That's a standard procedure, the minimax principle.Being kind of a developer myself (actually I'm mathematician) I didn't have a problem with that from the start, but I imagined upcoming confusion. There seems to be a principle: Everything that can be misunderstood will be misunderstood

 Author: djhbrown [ Thu Jan 04, 2018 7:25 am ] Post subject: Re: Video series: Learning from AlphaGo Baywa wrote:From a developer's point of view this makes a lot of sense.Yes, indeed, convenience for the programmer and les users peuvent manger de la brioche. Typical, bloody typical.One more for the Hall of ShameAttachment: hci.png [ 185.72 KiB | Viewed 3210 times ]

 Author: gennan [ Sat Jan 06, 2018 12:50 pm ] Post subject: Re: Video series: Learning from AlphaGo I just finished episode #3 about the Chinese Opening: https://www.youtube.com/watch?v=qv0gg0RJpZQ

 Author: gennan [ Sun Jan 21, 2018 12:56 pm ] Post subject: Re: Video series: Learning from AlphaGo I just finished episode #4: 5-space extensions are wronghttps://www.youtube.com/watch?v=OapPVew_stM

 Author: gennan [ Tue Feb 06, 2018 1:43 pm ] Post subject: Re: Video series: Learning from AlphaGo Learning from AlphaGo #5: refuting a greedy josekihttps://www.youtube.com/watch?v=5ouxL3G_tIo

 Author: daal [ Wed Feb 07, 2018 3:11 am ] Post subject: Re: Video series: Learning from AlphaGo I learned something from your last video that I hadn't previously understood. It is your explanation starting around 3:35 pointing out that the kikashi doesn't lose anything and can be treated lightly, because even if black swallows up the white stone, the loss is compensated by the fact that b had been forced to play inside his own territory. Thanks! Keep up the good work, I like your videos a lot. BTW, why does b always have such a miserable winrate, and is this opening winrate of 47% reflected in the overall results of alphago vs. alphago? If so, it seems a clear indication that komi is wrong, no?

 Author: luigi [ Wed Feb 07, 2018 5:38 am ] Post subject: Re: Video series: Learning from AlphaGo daal wrote:BTW, why does b always have such a miserable winrate, and is this opening winrate of 47% reflected in the overall results of alphago vs. alphago? If so, it seems a clear indication that komi is wrong, no?AlphaGo uses 7.5 komi and Chinese rules. 7 komi (breaking ties with a button) would probably be ideal.

 Author: Uberdude [ Wed Feb 07, 2018 5:53 am ] Post subject: Re: Video series: Learning from AlphaGo Yes, AG thinks white is better on the empty board with 7.5 komi. But I don't think that 47% number actually means black wins 47% of the million self-play games they've done, I understand it as an ill-defined goodness metric rather than a real probabilty, see discussion.If komi was 6.5 (does that even make sense with Chinese, does it need to change in 2s?) it could be black's "win rate" is 55%, so more lopsided than with 7.5; thus 7.5 might be the non-drawing komi which is closest to even which is probably what you want*. * Unless you go crazy and make a virtual fractional komi: say if 6.5 komi black wins 55% (+5%) and 7.5 komi 47% (-3%), therefore let's have 7 komi and if it's a draw on the board we assign win to black in 3/8 of cases and white in 5/8 of cases so komi is kinda like 7 and one eighth, where the 50% would be on a linear interpolation between our 2 data points.

 Author: djhbrown [ Wed Feb 07, 2018 5:59 am ] Post subject: Re: Video series: Learning from AlphaGo luigi wrote:[...komiheading off-topic here, so let's take it there:viewtopic.php?p=227322#p227322

 Author: luigi [ Wed Feb 07, 2018 6:33 am ] Post subject: Re: Video series: Learning from AlphaGo Uberdude wrote:7 komi and if it's a draw on the board we assign win to black in 3/8 of cases and white in 5/8 of cases so komi is kinda like 7 and one eighthWell, with the button, if it's a draw on the board, each color is assigned the win in 1/2 of cases.

 Author: gennan [ Wed Feb 07, 2018 12:27 pm ] Post subject: Re: Video series: Learning from AlphaGo It seems that 47% winrate for black with an empty board at 7.5 komi is even optimistic. From the 50 published self-play games Master vs Master, black only won 12 (24%). Also, in AlphaGo's opening database, you gradually see black's winrate drop as the game progresses, even when black plays the best moves according to AlphaGo. Usually black's winrate drops to about 43% by move 30. So as the game progresses, AlphaGo seems to become more certain that 7.5 komi is too much.I think that mathematically perfect komi should be an integer. And because of the increasingly bad odds that black seems to have in AlghaGo's evaluation and self-play game results (using 7.5 komi), my guess would be that perfect komi is 6 rather than 7. But for human (imperfect) games, I prefer a komi that prevents jigo. So perhaps 6.5 komi would be a fair approximation with Japanese rules. With Chinese rules, almost all games end with an odd score difference on the board. So in practice, black needs 7 points more on the board to win with 5.5, 6 or 6.5 komi. But black needs 9 points more on the board to win with 7.5 komi. (note that even though even score differences on the board are rare with Chinese rules, it is possible that the perfect game has it).

 Author: luigi [ Wed Feb 07, 2018 1:04 pm ] Post subject: Re: Video series: Learning from AlphaGo gennan wrote:It seems that 47% winrate for black with an empty board at 7.5 komi is even optimistic. From the 50 published self-play games Master vs Master, black only won 12 (24%). Also, in AlphaGo's opening database, you gradually see black's winrate drop as the game progresses, even when black plays the best moves according to AlphaGo. Usually black's winrate drops to about 43% by move 30. So as the game progresses, AlphaGo seems to become more certain that 7.5 komi is too much.That's only natural. If AlphaGo played perfectly, it would only display 0% and 100% win rates. It gets closer to perfection as the game progresses (of course, it's never going to display 50% right before passing).Also, the 50 published games are cherry-picked.

 Author: gennan [ Wed Feb 07, 2018 1:49 pm ] Post subject: Re: Video series: Learning from AlphaGo luigi wrote:gennan wrote:It seems that 47% winrate for black with an empty board at 7.5 komi is even optimistic. From the 50 published self-play games Master vs Master, black only won 12 (24%). Also, in AlphaGo's opening database, you gradually see black's winrate drop as the game progresses, even when black plays the best moves according to AlphaGo. Usually black's winrate drops to about 43% by move 30. So as the game progresses, AlphaGo seems to become more certain that 7.5 komi is too much.That's only natural. If AlphaGo played perfectly, it would only display 0% and 100% win rates. It gets closer to perfection as the game progresses (of course, it's never going to display 50% right before passing).Also, the 50 published games are cherry-picked.Do you think that DeepMind cherry-picked mostly white wins, but in fact unpublished wins had an even color distibution?Suppose that AlphaGo became so strong that it would evaluate 0% or 100% black winrate for an empty board with 7.5 komi. Will it be 0% or 100%? My guess is it would be 0%. But with an integer komi, there can be a komi value where perfect players would always get a jigo. The correct evaluation of an empty board with that komi value would be 50%.

 Page 2 of 4 All times are UTC - 8 hours [ DST ] Powered by phpBB © 2000, 2002, 2005, 2007 phpBB Grouphttp://www.phpbb.com/