symmetrical english is pretty solid. fiachetto bishop. and if possible get dr4 in at some point if not possible then prepare c5 is usual. Any openign has plenty of variations not matching the original intent. with englidh you may end up situation where besit continuation is to switch queens gambit. Or Panov attack. but quire often I get arrange my pieces and looks where to go forward
@petri999 said in #11:
> symmetrical english is pretty solid. fiachetto bishop. and if possible get dr4 in at some point if not possible then prepare c5 is usual. Any openign has plenty of variations not matching the original intent. with englidh you may end up situation where besit continuation is to switch queens gambit. Or Panov attack. but quire often I get arrange my pieces and looks where to go forward
How would you plan to go forward after these first 6 symetrical moves, as White? Stockfish suggests pawn to d4 immediately, but then I feel differently about it than stockfish, looking to snowball the advantage. You are right, symetrical English seems to be pretty solid. And, I think it is frustrating to play against. Also, it appears the move Nf3 when reflected with Nf6 gave White a slight evaluation increase, but I don't understand why either. It's like Nf3 prepares d4 for White, but Nf6 doesn't prepare d5 for Black?
r1bq1rk1/pp1pppbp/2n2np1/2p5/2P5/2N2NP1/PP1PPPBP/R1BQ1RK1 w - - 6 7
> symmetrical english is pretty solid. fiachetto bishop. and if possible get dr4 in at some point if not possible then prepare c5 is usual. Any openign has plenty of variations not matching the original intent. with englidh you may end up situation where besit continuation is to switch queens gambit. Or Panov attack. but quire often I get arrange my pieces and looks where to go forward
How would you plan to go forward after these first 6 symetrical moves, as White? Stockfish suggests pawn to d4 immediately, but then I feel differently about it than stockfish, looking to snowball the advantage. You are right, symetrical English seems to be pretty solid. And, I think it is frustrating to play against. Also, it appears the move Nf3 when reflected with Nf6 gave White a slight evaluation increase, but I don't understand why either. It's like Nf3 prepares d4 for White, but Nf6 doesn't prepare d5 for Black?
r1bq1rk1/pp1pppbp/2n2np1/2p5/2P5/2N2NP1/PP1PPPBP/R1BQ1RK1 w - - 6 7
Depth 55 is holey as Swiss cheese. The computer only consuders a tiny perventage of possible moves.
Play whatever you like. You'll win if you know it better than the other person.
Play whatever you like. You'll win if you know it better than the other person.
@Lojique said in #5:
> You're both right, this in and of itself is not really anything major or game-changing, I just found it kind of interesting so I thought I'd check myself to see how far I could get and report what it said. For those curious on what opening it came up with beyond just e4, the result was the Spanish Opening.
does more depth mean more accurate? maybe in a complete legal tree exploration with an accurate leaf node evaluation on un biased selection of leaf status.
Story (as i understand it now, burden of proof on new information, that must be quite recent...)
Horizon effect was fixed by adding a selective bias of leaf, it had to be palatable to a material signal dominated evaluation function at the time, because otherwise a leaf could be in the middle of a sequence of material exchange and the positional signal which could have predicted the material story after, were not part of the information visible by the eval paramétrisation choices). Not much has changed since then. iterative deepening, strict AB pruning, and i don't know when before SF11, non-strict AB pruning using forward pruning, in the iterative depth direction (not just the move pre-ordering one).
So a bet on depth before width. And a parameter optimization of an evolving eval function always conditional to biased set of input position (never learning to evaluate the positional signals in other positions).
so i ask again. is depth always giving a better variation? can it always recuperate branches that were dismissed before by some other path not yet dismissed, but explored deeper? in other words? yes, seem to the reflex answer. but i am not sure.
> You're both right, this in and of itself is not really anything major or game-changing, I just found it kind of interesting so I thought I'd check myself to see how far I could get and report what it said. For those curious on what opening it came up with beyond just e4, the result was the Spanish Opening.
does more depth mean more accurate? maybe in a complete legal tree exploration with an accurate leaf node evaluation on un biased selection of leaf status.
Story (as i understand it now, burden of proof on new information, that must be quite recent...)
Horizon effect was fixed by adding a selective bias of leaf, it had to be palatable to a material signal dominated evaluation function at the time, because otherwise a leaf could be in the middle of a sequence of material exchange and the positional signal which could have predicted the material story after, were not part of the information visible by the eval paramétrisation choices). Not much has changed since then. iterative deepening, strict AB pruning, and i don't know when before SF11, non-strict AB pruning using forward pruning, in the iterative depth direction (not just the move pre-ordering one).
So a bet on depth before width. And a parameter optimization of an evolving eval function always conditional to biased set of input position (never learning to evaluate the positional signals in other positions).
so i ask again. is depth always giving a better variation? can it always recuperate branches that were dismissed before by some other path not yet dismissed, but explored deeper? in other words? yes, seem to the reflex answer. but i am not sure.
@dboing said in #14:
> does more depth mean more accurate? maybe in a complete legal tree exploration with an accurate leaf node evaluation on un biased selection of leaf status.
>
> Story (as i understand it now, burden of proof on new information, that must be quite recent...)
>
> Horizon effect was fixed by adding a selective bias of leaf, it had to be palatable to a material signal dominated evaluation function at the time, because otherwise a leaf could be in the middle of a sequence of material exchange and the positional signal which could have predicted the material story after, were not part of the information visible by the eval paramétrisation choices). Not much has changed since then. iterative deepening, strict AB pruning, and i don't know when before SF11, non-strict AB pruning using forward pruning, in the iterative depth direction (not just the move pre-ordering one).
> So a bet on depth before width. And a parameter optimization of an evolving eval function always conditional to biased set of input position (never learning to evaluate the positional signals in other positions).
>
> so i ask again. is depth always giving a better variation? can it always recuperate branches that were dismissed before by some other path not yet dismissed, but explored deeper? in other words? yes, seem to the reflex answer. but i am not sure.
To make an illustration: Looking at a good photograph can give a good evaluation of Where's Waldo (checkmate) at Depth = 2.
However, looking at a good photograph more deeply - not being able to see Waldo at Depth = 2, we take "micro-scopes" of the 400 positions of Depth = 2 to more accurately evaluate the starting position.
And then, considering the best possibilities first with the next best possibilities of Where Waldo (checkmate) might be, given better evaluation scores, take priority over considering less likely moves in depth. So, some of the 400 positions get super-microscoped, whereas, other position hardly get micro-scoped at all, it would seem.
So, when the micro-scopping technique fails - that is the Horizon effect in a nutshell, because it didn't complete the analysis of tacitics in a tactical position. One side may have been materially ahead temporarily, but is in a totally losing position.
Therefore, how deep of micro-scoping the position of the 400 key positions reliable, until we need a better way of evaluating the position?
A chess position that is 99% evaluated perfectly still, when micro-scopped (for accuracy) still gets it wrong at a compounded interest rate. 1% inaccurate becomes (100% - 99% of 99%) with 1 microscope, but adds confidence to each move also by a compounded rate too.
What is the equilibrium of micro-scopes compared to having a better 1st picture to take microscopes from?
Yes, deeper is generally better than wider, to a point.
I wonder if there are any mathematics done on this to enhance Stockfish's performance?
> does more depth mean more accurate? maybe in a complete legal tree exploration with an accurate leaf node evaluation on un biased selection of leaf status.
>
> Story (as i understand it now, burden of proof on new information, that must be quite recent...)
>
> Horizon effect was fixed by adding a selective bias of leaf, it had to be palatable to a material signal dominated evaluation function at the time, because otherwise a leaf could be in the middle of a sequence of material exchange and the positional signal which could have predicted the material story after, were not part of the information visible by the eval paramétrisation choices). Not much has changed since then. iterative deepening, strict AB pruning, and i don't know when before SF11, non-strict AB pruning using forward pruning, in the iterative depth direction (not just the move pre-ordering one).
> So a bet on depth before width. And a parameter optimization of an evolving eval function always conditional to biased set of input position (never learning to evaluate the positional signals in other positions).
>
> so i ask again. is depth always giving a better variation? can it always recuperate branches that were dismissed before by some other path not yet dismissed, but explored deeper? in other words? yes, seem to the reflex answer. but i am not sure.
To make an illustration: Looking at a good photograph can give a good evaluation of Where's Waldo (checkmate) at Depth = 2.
However, looking at a good photograph more deeply - not being able to see Waldo at Depth = 2, we take "micro-scopes" of the 400 positions of Depth = 2 to more accurately evaluate the starting position.
And then, considering the best possibilities first with the next best possibilities of Where Waldo (checkmate) might be, given better evaluation scores, take priority over considering less likely moves in depth. So, some of the 400 positions get super-microscoped, whereas, other position hardly get micro-scoped at all, it would seem.
So, when the micro-scopping technique fails - that is the Horizon effect in a nutshell, because it didn't complete the analysis of tacitics in a tactical position. One side may have been materially ahead temporarily, but is in a totally losing position.
Therefore, how deep of micro-scoping the position of the 400 key positions reliable, until we need a better way of evaluating the position?
A chess position that is 99% evaluated perfectly still, when micro-scopped (for accuracy) still gets it wrong at a compounded interest rate. 1% inaccurate becomes (100% - 99% of 99%) with 1 microscope, but adds confidence to each move also by a compounded rate too.
What is the equilibrium of micro-scopes compared to having a better 1st picture to take microscopes from?
Yes, deeper is generally better than wider, to a point.
I wonder if there are any mathematics done on this to enhance Stockfish's performance?
<Comment deleted by user>
@Lojique said in #1:
> After running the Stockfish analysis for several hours on a spare machine I had I've found that at Depth 55 the best first move for white becomes e4 again; currently the cloud says that it is Nf3 based on a depth of 53. I was able to get to Depth 55 twice and it said e4 both times, although the first time the page refreshed itself before I could take a screenshot (probably because I was away for too long since I left it idle running on its own the whole time), but the second time I was able to take a screenshot before it refreshed so I actually can confirm it if necessary.
e4, d4, c4, f4, nf3, b3, g3 are all playable.
> After running the Stockfish analysis for several hours on a spare machine I had I've found that at Depth 55 the best first move for white becomes e4 again; currently the cloud says that it is Nf3 based on a depth of 53. I was able to get to Depth 55 twice and it said e4 both times, although the first time the page refreshed itself before I could take a screenshot (probably because I was away for too long since I left it idle running on its own the whole time), but the second time I was able to take a screenshot before it refreshed so I actually can confirm it if necessary.
e4, d4, c4, f4, nf3, b3, g3 are all playable.
@Approximation said in #12:
> How would you plan to go forward after these first 6 symetrical moves, as White? Stockfish suggests pawn to d4 immediately,
> r1bq1rk1/pp1pppbp/2n2np1/2p5/2P5/2N2NP1/PP1PPPBP/R1BQ1RK1 w - - 6 7
I would definitely play d4 here. I want remove black control of centre. while it is possible. I think I have occasionally played d3 and after that a3 b4. but d4 time is about now before black comes to pin your Kn
> How would you plan to go forward after these first 6 symetrical moves, as White? Stockfish suggests pawn to d4 immediately,
> r1bq1rk1/pp1pppbp/2n2np1/2p5/2P5/2N2NP1/PP1PPPBP/R1BQ1RK1 w - - 6 7
I would definitely play d4 here. I want remove black control of centre. while it is possible. I think I have occasionally played d3 and after that a3 b4. but d4 time is about now before black comes to pin your Kn
interesting
AlphaZero is still laughing.
This topic has been archived and can no longer be replied to.