Alpha Zero's 21. Bg5! revisited

Discussion of anything and everything relating to chess playing software and machines.

Moderators: bob, hgm, Harvey Williamson

Forum rules
This textbox is used to restore diagrams posted with the [d] tag before the upgrade.
mmt
Posts: 274
Joined: Sun Aug 25, 2019 6:33 am
Full name: .

Re: Alpha Zero's 21. Bg5! revisited

Post by mmt » Wed Sep 16, 2020 6:51 pm

Vinvin wrote:
Wed Sep 16, 2020 6:23 pm
I remember I spent some time, some years ago on this position and Bg5 was consistently 1 pawn above b4.
Try to run with 2 best moves or see the score after playing both moves.
I got the position after Bg5 evaluated up the depth of 65 now and it indeed has a higher score by 1 pawn (-4.59). But since it seems very likely that 21. b4 is also winning, it's not a good test position. This was also noticed earlier here viewtopic.php?t=68955 though I analyzed it much deeper.

User avatar
MikeB
Posts: 4360
Joined: Thu Mar 09, 2006 5:34 am
Location: Pen Argyl, Pennsylvania

Re: Alpha Zero's 21. Bg5! revisited

Post by MikeB » Thu Sep 17, 2020 1:01 am

Vinvin wrote:
Wed Sep 16, 2020 6:23 pm
mmt wrote:
Wed Sep 16, 2020 1:45 pm
Do we actually know that 21. Bg5 is better?
I remember I spent some time, some years ago on this position and Bg5 was consistently 1 pawn above b4.
Try to run with 2 best moves or see the score after playing both moves.
Correct, Bg5 is the best move. Stockfsh will show about a pawn, I dampen the score output for honey engines by about 1/3 from SF - just my own view that SF score is a little high , a little too optimistic, ymmv

Bluefish
info depth 37 seldepth 74 multipv 1 score cp 218 nodes 37057673397 nps 73906634 hashfull 989 tbhits 9770 time 501412 pv c1g5 f7f5 g4f4 d7d5 e4f6 g7f6 g5f6 f8f6 e1e8 h7g8 e8g8 h8g8 f4c7 f6f7 c7c8 g8g7 a1e1 b7c5 c8d8 c5e4 f2f3 a6b7 f3e4 b8d7 d8c7 d7f6 c7e5 f5e4 e5f5 b7c8 f5g6 g7f8 g6h6 f8e7 h6g5 c8e6 e1f1 a8f8 h5h6 e7d6 g3g4 f6h7 g5h4 f7f1 g2f1 d6d7 f1e2 f8g8 h4g3 h7g5 g3e5 g8g6 e5h2 g5h7 h2b8 e6g4 e2g4 g6g4 g1f1 h7f6
info depth 36 seldepth 60 multipv 2 score cp 158 nodes 37057673397 nps 73906634 hashfull 989 tbhits 9770 time 501412 pv b2b4 d7d5 e4d2 a6d3 d2f3 b7d6 c1b2 b8a6 f3e5 d3e4 e5c6 a6c7 b2e5 c7e8 a1c1 e4g2 g1g2 d6e4 g4f4 e8f6 f2f3 e4g5 f4f6 g7f6 e5f6 h7g7 f6g7 h8g7 e1d1 f8c8 d1d5 c8c7 c1c2 a8e8 d5d2 e8e3 c6d4 c7c2 d2c2 e3d3 d4f5 g7h7 c2c7 d3d2 g2f1 g5f3 c7f7 h7h8 f5h6 f3h2 f1g1 d2a2 f7f4 a7a5 b4a5 b6a5
Image

mmt
Posts: 274
Joined: Sun Aug 25, 2019 6:33 am
Full name: .

Re: Alpha Zero's 21. Bg5! revisited

Post by mmt » Sat Sep 19, 2020 4:24 pm

MikeB wrote:
Thu Sep 17, 2020 1:01 am
Correct, Bg5 is the best move.
Well, since b4 is +3.59, 99.9% chance they both lead to a win. So it's not really THE best move. It is just one of at least two winning moves.

peter
Posts: 2132
Joined: Sat Feb 16, 2008 6:38 am
Full name: Peter Martan

Re: Alpha Zero's 21. Bg5! revisited

Post by peter » Sat Sep 19, 2020 6:48 pm

mmt wrote:
Sat Sep 19, 2020 4:24 pm
MikeB wrote:
Thu Sep 17, 2020 1:01 am
Correct, Bg5 is the best move.
Well, since b4 is +3.59, 99.9% chance they both lead to a win. So it's not really THE best move. It is just one of at least two winning moves.
Well, even if there were more then two moves leading to a win, there must be always one of them the best one. My bets are on Bg5!.

Sophisticated regards
Peter.

mmt
Posts: 274
Joined: Sun Aug 25, 2019 6:33 am
Full name: .

Re: Alpha Zero's 21. Bg5! revisited

Post by mmt » Sat Sep 19, 2020 6:51 pm

peter wrote:
Sat Sep 19, 2020 6:48 pm
Well, even if there were more then two moves leading to a win, there must be always one of them the best one. My bets are on Bg5!.
Why? A mate is a mate, doesn't matter in how many moves.

peter
Posts: 2132
Joined: Sat Feb 16, 2008 6:38 am
Full name: Peter Martan

Re: Alpha Zero's 21. Bg5! revisited

Post by peter » Sat Sep 19, 2020 6:52 pm

mmt wrote:
Sat Sep 19, 2020 6:51 pm
peter wrote:
Sat Sep 19, 2020 6:48 pm
Well, even if there were more then two moves leading to a win, there must be always one of them the best one. My bets are on Bg5!.
Why? A mate is a mate, doesn't matter in how many moves.
You don't like mate in x- problems, do you?
And I know this one isn't one, but if you aren't able to find the mating distance for sure, won't you try to have other ways to dinstinct a good and probably winning move from a better one move winning with even more probability?
And if you think a move will lead to a sooner mate then another one probably winning too, which one would you choose?
I mean, you'll have to choose a single one move always, not being able to make two moves at once and win a single game in that way twice, haven't you to?
:)
Nevermind regards
Peter.

mmt
Posts: 274
Joined: Sun Aug 25, 2019 6:33 am
Full name: .

Re: Alpha Zero's 21. Bg5! revisited

Post by mmt » Sat Sep 19, 2020 7:16 pm

peter wrote:
Sat Sep 19, 2020 6:52 pm
You don't like mate in x- problems, do you?
And I know this one isn't one, but if you aren't able to find the mating distance for sure, won't you try to have other ways to dinstinct a good and probably winning move from a better one move winning with even more probability?
And if you think a move will lead to a sooner mate then another one probably winning too, which one would you choose?
I mean, you'll have to choose a single one move always, not being able to make two moves at once and win a single game in that way twice, haven't you to?
:)
Nevermind regards
I actually wrote a program to solve mate in X problems many years ago :)

I mean I agree that Bg5 is a prettier move of course but my point is that all the judgments about whether one version of SF or LC0 is better than another or as good as AlphaZero based on this position are very likely wrong. If a program chooses b4, it might be able to "understand" this position just as well as another program that plays Bg5.

peter
Posts: 2132
Joined: Sat Feb 16, 2008 6:38 am
Full name: Peter Martan

Re: Alpha Zero's 21. Bg5! revisited

Post by peter » Sat Sep 19, 2020 7:28 pm

mmt wrote:
Sat Sep 19, 2020 7:16 pm
peter wrote:
Sat Sep 19, 2020 6:52 pm
You don't like mate in x- problems, do you?
And I know this one isn't one, but if you aren't able to find the mating distance for sure, won't you try to have other ways to dinstinct a good and probably winning move from a better one move winning with even more probability?
And if you think a move will lead to a sooner mate then another one probably winning too, which one would you choose?
I mean, you'll have to choose a single one move always, not being able to make two moves at once and win a single game in that way twice, haven't you to?
:)
Nevermind regards
I actually wrote a program to solve mate in X problems many years ago :)

I mean I agree that Bg5 is a prettier move of course but my point is that all the judgments about whether one version of SF or LC0 is better than another or as good as AlphaZero based on this position are very likely wrong. If a program chooses b4, it might be able to "understand" this position just as well as another program that plays Bg5.
So you say b4 is better?
I mean, if you have to choose as a chess player, having seen many good lines after both moves, forgetting about "test positions" and their very many all subjective definitions of them.

Want to play b4 with White in a corr. game?
In this case you'd have to have a second one against Bg5 with Black too, which one move would you choose in this case?

Ready for the rumble regards
:)
Peter.

mmt
Posts: 274
Joined: Sun Aug 25, 2019 6:33 am
Full name: .

Re: Alpha Zero's 21. Bg5! revisited

Post by mmt » Sat Sep 19, 2020 7:47 pm

peter wrote:
Sat Sep 19, 2020 7:28 pm
So you say b4 is better?
No, I say that they are very likely equal.
peter wrote:
Sat Sep 19, 2020 7:28 pm
I mean, if you have to choose as a chess player, having seen many good lines after both moves, forgetting about "test positions" and their very many all subjective definitions of them.

Want to play b4 with White in a corr. game?
In this case you'd have to have a second one against Bg5 with Black too, which one move would you choose in this case?
I'd play a prettier move, even if it was slightly worse, I'm not playing for money. It doesn't really matter for determining which move is objectively better, though.

If you have a position where the supposed winning move is +4.6 but there is another with +3.6, it's just not a useful test position.

peter
Posts: 2132
Joined: Sat Feb 16, 2008 6:38 am
Full name: Peter Martan

Re: Alpha Zero's 21. Bg5! revisited

Post by peter » Sat Sep 19, 2020 8:07 pm

mmt wrote:
Sat Sep 19, 2020 7:47 pm
peter wrote:
Sat Sep 19, 2020 7:28 pm
So you say b4 is better?
No, I say that they are very likely equal.
peter wrote:
Sat Sep 19, 2020 7:28 pm
I mean, if you have to choose as a chess player, having seen many good lines after both moves, forgetting about "test positions" and their very many all subjective definitions of them.

Want to play b4 with White in a corr. game?
In this case you'd have to have a second one against Bg5 with Black too, which one move would you choose in this case?
I'd play a prettier move, even if it was slightly worse, I'm not playing for money. It doesn't really matter for determining which move is objectively better, though.
Which way is a good one for you to determine one move winning is better then another one probably winning too?

I you aren't able to discriminate one move to be better then an other one, than that's not a good (tactical) test position, as long as you don't give same amount of points for one, two, or even three moves you consider all of just the same quality and to all other moves, you judge to be clearly worse, no points at all or less points, in that way you can of course even such positions use for (e.g. positional) testing. It's what you do all the time, choosing opening positions for eng-eng-matches, don't you?

So let's stop this rather academic discussion (which for me is an ironical synonyme for discussions of little pracitical meaning) here. You just have a narrower definition of test position than I have, and Bg5 is better then b4 yet still, isn't it?
I mean, why choose a move with eval 3.6, if you have a move with eval 4.6, which, if I remember your postings correctly, are the evals you trust in?
If you don't trust in any engine- evals at all, why quote them (no matter whether given in centipawn, wdl- probability or whatsoever)?

No more rumble from my side regards
:)
Peter.

Post Reply