Page 2 of 2
Re: AI vs Dummies - Dummies win?
Posted: Tue Feb 05, 2019 10:33 am
by Aidoneus
Bill Spight wrote:jlt wrote:Playing locally is 24 points gote. At the opening stage, it is possible that some tenuki moves are worth more than that.
By swing counting, the first move on an empty board is worth approximately 28 pts.
Except the first move isn't worth a 28-point komi.
Re: AI vs Dummies - Dummies win?
Posted: Tue Feb 05, 2019 10:37 am
by jlt
Here is a totally artificial board position where LZ157 thinks that tenuki is better.
- Capture.JPG (93.76 KiB) Viewed 13867 times
Re: AI vs Dummies - Dummies win?
Posted: Tue Feb 05, 2019 10:51 am
by jlt
Aidoneus wrote:Except the first move isn't worth a 28-point komi.
Probably Bill has a better better explanation but here is how I interpret this value of 28 points. Suppose perfect komi is 7.
Consider a simplified model where move 1 is a points, move 2 is a-h points, move 3 is a-2h points,... until move 2n which is a-(2n-1)h points.
There is no extra move available, so a=2nh. The difference of points between Black and White is nh, so nh=7 and a=14.
Swing counting consists in adding points by Black if Black makes one move, and points by White if White makes one move, so we get a+a=28 points.
Re: AI vs Dummies - Dummies win?
Posted: Tue Feb 05, 2019 10:57 am
by Aidoneus
Sure, I did understand that we are looking at an alternating sequence, which is why I don't believe tenuki would be worthwhile on an EMPTY board. But I guess my wink was wasted, yes?
Re: AI vs Dummies - Dummies win?
Posted: Tue Feb 05, 2019 11:10 am
by jlt
Your last message seems to imply that you disagree with Bill, but you don't really explain why.
Re: AI vs Dummies - Dummies win?
Posted: Tue Feb 05, 2019 11:14 am
by Aidoneus
jlt wrote:Here is a totally artificial board position where LZ157 thinks that tenuki is better.
IMHO, this is simply another example, like ladders, in which the AI isn't looking far enough ahead.
Re: AI vs Dummies - Dummies win?
Posted: Tue Feb 05, 2019 11:24 am
by Bill Spight
Here is a Deep Leela board to illustrate the point.
Explanation of the diagram:
The White 3-3 stone in the bottom right corner is there to minimize the possible necessity for White to kill the enclosed Black group on the bottom side in order to make its own group live.
The Black stones on the rest of the board are there to make the whole board, in Deep Leela's estimation approximately equal with Black to play, while leaving as much of the board open. In playing around with Deep Leela, I have discovered that it makes the best decisions, from a human point of view, when it thinks that the game is even. (I don't know about other bots.)
Given that, Deep Leela prefers Black A - D in order, with all of them worth around a 50% win rate. There are other possible moves in the top right corner and right side with approximately 50% win rates that I did not show. Black E is on Deep Leela's radar, while K-01 is not, but with a win rate estimate of around 40%.
Re: AI vs Dummies - Dummies win?
Posted: Tue Feb 05, 2019 11:31 am
by Bill Spight
Re: AI vs Dummies - Dummies win?
Posted: Tue Feb 05, 2019 11:41 am
by Aidoneus
Bill Spight wrote:
Explanation of the diagram:
Given that, Deep Leela prefers Black A - D in order, with all of them worth around a 50% win rate.
I am quite surprised that B is its second choice. Would you mind offering some of your dan-level insight?
Re: AI vs Dummies - Dummies win?
Posted: Tue Feb 05, 2019 12:17 pm
by Bill Spight
Aidoneus wrote:Bill Spight wrote:
Explanation of the diagram:
Given that, Deep Leela prefers Black A - D in order, with all of them worth around a 50% win rate.
I am quite surprised that B is its second choice. Would you mind offering some of your dan-level insight?
Sorry for not being clear. Deep Leela's second choice is actually the 4-4 in the top right. B is the second area that it likes. But all of the plays it suggests, except for killing the group, have estimated win rates of around 50%.
Re: AI vs Dummies - Dummies win?
Posted: Tue Feb 05, 2019 12:42 pm
by Aidoneus
Bill Spight wrote:Aidoneus wrote:Bill Spight wrote:
Explanation of the diagram:
Given that, Deep Leela prefers Black A - D in order, with all of them worth around a 50% win rate.
I am quite surprised that B is its second choice. Would you mind offering some of your dan-level insight?
Sorry for not being clear. Deep Leela's second choice is actually the 4-4 in the top right. B is the second area that it likes. But all of the plays it suggests, except for killing the group, have estimated win rates of around 50%.
I see, after some move in the top right, its second
area is a move on the side (B). Thank you for the clarification.
Re: AI vs Dummies - Dummies win?
Posted: Tue Feb 05, 2019 2:51 pm
by jaca
Bill Spight wrote:K-01
Dear Sir,
On behalf of the Campaign for United Columnar Kinship Open Operations, I wish to complain in the strongest possible terms about the unfair discrimination against our gender - why should Rowars get twice as much space as us?! We work just as hard as them to label jamjars, so we deserve equal representation.
Yours Apoplectically,
Lt.Col. Pankhurst, E., Mrs. (retd.),
Tunbridge Wells
----------------------------------
BS-PS
Aidoneus wrote:I am quite surprised that B is its second choice.
Dear Sir,
On behalf of the Dans Obeisance to Pippen Exultation Society, I wish to complain in the strongest possible terms about this Insidious Succurilous Implied Slight upon the Holy name of our Glorious Overlord's glorious Pippen-Fuseki (swt).
Of course the side is Bigger than the corner! Everyone knows that B is the Beautiful, Best, Biggest, Billion Bullion Bubble Bush Sr, Bush Jr, Boris Bus Bullocks,... sorry, i can't think of any other words Beginning with B.
Code: Select all
In the Beginning was the B, and the B was with Pippen, and the B was Pippen - Thou shalt have no other Go but Pippen-Fuseki - Pippen Uber Balles - Ballons, enfants de la Patrie de Pippen!
Re: AI vs Dummies - Dummies win?
Posted: Thu Feb 07, 2019 1:55 pm
by jaca
Aidoneus wrote: I don't believe tenuki
- seed.jpeg (22.57 KiB) Viewed 4528 times
- large shimari.png (130.1 KiB) Viewed 4521 times
Michael Redmond wrote: This large shimari wasn't played much until Alphago started playing it in the Master series
and now the tossers down my end of a back street do it all the time. Monkey see, monkey do.. but we are all simian, so it's maybe not a bad strategy, for all young children initially learn by imitation too.
So far, so good - by why did Alphago do it? And does LZ do it too? i can't answer the second question (my guess is yes), but i can have a stab at the first:
DCNN+feedback is all about probability, even if random rollouts are not involved, so moves that give opp the most chances of stuffing up become valued.
There are holes in the large shimari all over the place, providing dozens of opportunities for shooting yourself in the foot, and LZ can pretty well read them all out to 20 deep, so no wonder she likes it.
But does that mean it's a good idea for anyone who can't read 2 gigaploys in 3 microseconds?
Your grandfather didn't think so, and it wouldn't surprise me if even Iyama Yuta eventually comes around to agreeing with Shusaku.
