Uri Blass wrote:Most of the moves are not forced so this advantage is very small.
But it is magnified the more games you play simultaneously.
A player that assists the engine will be able to, say, play 24 games concurrently, at about the same strenght it'd play one, because it'd just take longer to move, and focus in the games where it has the lowest time remaining, to avoid losing any game on time.
How is the unassited engine going to achieve this? Seems it'd be able to barely play 2 games at a time.
Leo wrote:It seems like a waste when I play matches on the computer and use all that energy and when the game is over, all the calculations disappear.
I recommend Mike Leahy's Chess Openings Wizard for this.
Unfortunately, it's not as useful as you'd think. Sure, I have all my analysis from the past 10 years in a tidy form, the problem is that when I get to use it, it turns out it's not sound anymore, and I'm able to find much better lines with what is available today.
I don't even think my analysis from 6 months ago is useful anymore
So with analysis getting obsolete in record time, you might as well accept that no matter what, either the calculations disappear or they're not going to be useful in the future.
I use COW's !! symbol for the absolute best moves I've been able to find, that have taken me hours to develop, only to see an opponent play into that position, but after double checking it turns out the variations fall down and I can't rescue them, and the move gets assigned a !? and may become third best...
I wonder how will Cerebellum look 1 year from now, perhaps it'll have a completely different repertoire, as the strongest engines of the year get released and bust its current mainlines. Perhaps we will see the come back of 1.e4!
Here is what Cerebellums website says :
"The book is always calculated with the newest Version. When the tree search is being improved, there is no big problem, Stockfish basically only searches more efficient and faster. When the evaluation function changes, in theory you would have to recalculate the whole book. But because you do not calculate only new lines, but regularly already existing lines are extended too, the book only needs some time to adapt to the new evaluation. I would say around a week. So there are really positions in the book which where calculated with a 2 year old Stockfish, but that is no problem, because they are now "hidden" under a big amount of extended positions, which now dominate the calculation of the tree, because the new nodes are always closer to the terminating nodes of the tree."
Uri Blass wrote:Most of the moves are not forced so this advantage is very small.
But it is magnified the more games you play simultaneously.
A player that assists the engine will be able to, say, play 24 games concurrently, at about the same strenght it'd play one, because it'd just take longer to move, and focus in the games where it has the lowest time remaining, to avoid losing any game on time.
How is the unassited engine going to achieve this? Seems it'd be able to barely play 2 games at a time.
I do not understand.
Even if 60% of the moves that you play in games are not forced then you can reply immediately in 40% of the cases but need to use time in 60% of the cases.
You can use more time thanks to replying immediately in 40% of the cases but I do not see how this advantage is even 2:1 and you suggest that it is 24:1
"The book is always calculated with the newest Version. When the tree search is being improved, there is no big problem, Stockfish basically only searches more efficient and faster. When the evaluation function changes, in theory you would have to recalculate the whole book. But because you do not calculate only new lines, but regularly already existing lines are extended too, the book only needs some time to adapt to the new evaluation. I would say around a week. So there are really positions in the book which where calculated with a 2 year old Stockfish, but that is no problem, because they are now "hidden" under a big amount of extended positions, which now dominate the calculation of the tree, because the new nodes are always closer to the terminating nodes of the tree."
COW Books can adapt immediately by using the "EPD->Export Positions" and "EPD->Analyze with engine" features. You don't have to wait weeks, as you can decide if you only update new lines or if you update every node of the book.
For instance, I downloaded the PGNs from Carlos Canavessi's LS Elite League and Superfinal, I import them into a Bookup's Ebook up to 84 ply, then I "EPD->Export Positions" to get all leaf nodes into an analysis file, I start up Sugar xPRO as my engine (I use this because it automatically uses all my cores), I use "EPD->Analyze with engine", then I move to the root position and Backsolve...
I get scores for all lines. It will show that some moves are winning if there's no alternative move that avoids the losing position. I compare those lines with the ones from my Monster Book that already have deviations that don't lose and include it into the book...
BAM!
I have a book that includes all moves played up until 84 plies in each of the games of those tournaments and they have a score that the engine gave, that was backsolved to the root automatically.
And if Stockfish 9 is released, I could, in theory, use "EPD->Analyze with engine" (I don't even need to export the positions again, I already have them) and all the leaf nodes would be updated with Stockfish scores, not only new lines.
Now, I feel awkward to be promoting what is basically Cerebellum's competition in a Cerebellum's thread, but perhaps it is important to know that fully capable solutions already exists, with the difference that with COW you have to start from scratch, while Cerebellum already has its book full of mainlines, but it would be possible to turn Cerebellum into a COW book and have updated scores on the fly without waiting for new lines to arrive.
Even if 60% of the moves that you play in games are not forced then you can reply immediately in 40% of the cases but need to use time in 60% of the cases.
You can use more time thanks to replying immediately in 40% of the cases but I do not see how this advantage is even 2:1 and you suggest that it is 24:1
Because you only need the 24 hours for critical positions, the ones that decide games, the ones that are playing a losing move that you suggested to analyze with an engine for 24 hours.
The rest of the positions either need 0 time (forced move) or you don't need much time to decide because there's no correct move to find, and it's like deciding between 1.e4, 1.d4 or 1.Nf3, all moves are fine, you can see this in 10 minutes, make a choice and move on to the next game. The rest of the time, which is huge, can be used to analyze the critical positions much more deeply, and as this accumulate, you may analyze them even deeper than the computer maxed out at 24 hours/move. With this strategy you can match the unattended computer's strength and surpass it for the positions of the games that matter.
Of course I left out the details that this person needs to sleep, so for 8 hours of sleep you may decrease the number of opponents from 24 to 16.
Ovyron wrote:
Now, I feel awkward to be promoting what is basically Cerebellum's competition in a Cerebellum's thread, but perhaps it is important to know that fully capable solutions already exists, with the difference that with COW you have to start from scratch, while Cerebellum already has its book full of mainlines, but it would be possible to turn Cerebellum into a COW book and have updated scores on the fly without waiting for new lines to arrive.
I'm not sure where you see an advantage of this COW book. I have several million leave nodes in the book. Of course I can update them all on the fly, but it would take over a year to do this. Manual updates of nodes are always possible in Cerebellum, so you can always update and extend positions in which you are interested or for example which you have in an Coresspondence chess game.
Also the tree calculation is different. Cerebellum does not simply propagate values from the leave nodes to the root, it recalculates the whole database in a graph model, which is complicated but the only way to achieve correct results, because chess is a graph, not a tree.
Also there are very refined methods in Cerebellum to improve the calculation, I will describe them when the full Version appears.
Thomas Zipproth wrote:I'm not sure where you see an advantage of this COW book.
The advantage of COW is that people can buy it and use it already, nothing more. I was talking to people that can't wait for Cerebellum's release because they want to do some things that they might already be able to do.
Cerebellum does not simply propagate values from the leave nodes to the root, it recalculates the whole database in a graph model, which is complicated but the only way to achieve correct results, because chess is a graph, not a tree.
Can you explain the differenence?
If it's what I suspect, then, indeed, the backsolved scores I'm showing aren't right, if there was some possible transposition that was missing, that would link a position with high score to a refutation of a low score, and fixing this would require visiting every single node and doing the "Add Moves that transpose" thing, which is unfeasible.
Anyway, I didn't know you were dealing with several millions of leaf nodes, probably the size of the project makes it nearly impossible to export it outside Cerebellum.
Even if 60% of the moves that you play in games are not forced then you can reply immediately in 40% of the cases but need to use time in 60% of the cases.
You can use more time thanks to replying immediately in 40% of the cases but I do not see how this advantage is even 2:1 and you suggest that it is 24:1
Because you only need the 24 hours for critical positions, the ones that decide games, the ones that are playing a losing move that you suggested to analyze with an engine for 24 hours.
The rest of the positions either need 0 time (forced move) or you don't need much time to decide because there's no correct move to find, and it's like deciding between 1.e4, 1.d4 or 1.Nf3, all moves are fine, you can see this in 10 minutes, make a choice and move on to the next game. The rest of the time, which is huge, can be used to analyze the critical positions much more deeply, and as this accumulate, you may analyze them even deeper than the computer maxed out at 24 hours/move. With this strategy you can match the unattended computer's strength and surpass it for the positions of the games that matter.
Of course I left out the details that this person needs to sleep, so for 8 hours of sleep you may decrease the number of opponents from 24 to 16.
I do not agree and I think that you need also to decide between different moves that draw about the move that gives more practical chances to win.
In many games it is not clear what is the losing mistake and one side got positional advantage and increase it.
Without using the computer for a long time the winner could play worse moves and not get positional advantage.
I suggested to use old games to analyze because of the idea that with today hardware and stockfish you can find the losing mistake but you could not do it at that time and at that time using the computer for a long time could only show that the losing mistake is a positional mistake that lose 0.2 pawn that you do not see in a short analysis.
"The book is always calculated with the newest Version. When the tree search is being improved, there is no big problem, Stockfish basically only searches more efficient and faster. When the evaluation function changes, in theory you would have to recalculate the whole book. But because you do not calculate only new lines, but regularly already existing lines are extended too, the book only needs some time to adapt to the new evaluation. I would say around a week. So there are really positions in the book which where calculated with a 2 year old Stockfish, but that is no problem, because they are now "hidden" under a big amount of extended positions, which now dominate the calculation of the tree, because the new nodes are always closer to the terminating nodes of the tree."
COW Books can adapt immediately by using the "EPD->Export Positions" and "EPD->Analyze with engine" features. You don't have to wait weeks, as you can decide if you only update new lines or if you update every node of the book.
For instance, I downloaded the PGNs from Carlos Canavessi's LS Elite League and Superfinal, I import them into a Bookup's Ebook up to 84 ply, then I "EPD->Export Positions" to get all leaf nodes into an analysis file, I start up Sugar xPRO as my engine (I use this because it automatically uses all my cores), I use "EPD->Analyze with engine", then I move to the root position and Backsolve...
I get scores for all lines. It will show that some moves are winning if there's no alternative move that avoids the losing position. I compare those lines with the ones from my Monster Book that already have deviations that don't lose and include it into the book...
BAM!
I have a book that includes all moves played up until 84 plies in each of the games of those tournaments and they have a score that the engine gave, that was backsolved to the root automatically.
And if Stockfish 9 is released, I could, in theory, use "EPD->Analyze with engine" (I don't even need to export the positions again, I already have them) and all the leaf nodes would be updated with Stockfish scores, not only new lines.
Now, I feel awkward to be promoting what is basically Cerebellum's competition in a Cerebellum's thread, but perhaps it is important to know that fully capable solutions already exists, with the difference that with COW you have to start from scratch, while Cerebellum already has its book full of mainlines, but it would be possible to turn Cerebellum into a COW book and have updated scores on the fly without waiting for new lines to arrive.
This all seems incredibly interesting. I had never heard of Mike Leahy's Chess Openings Wizard. It will take me some time to investigate. It sounds like you already have something like cerebellum. The 2 together would maybe be even more powerful.