Dann Corbit wrote: ↑Wed Oct 30, 2019 4:08 pm
I recall a time when a chess engine would think all night to achieve 17 plies.
How would we have defined tactical and strategic then?
I think the definition is a moving target.
But I would put it this way:
A strategic move is made using first principles without seeing a definite win of material or game outcome.
A tactical move is made for a known purpose, either a win in material or mate or draw achieved.
Hence, for instance, giving up a full piece might be either tactical or strategic.
So who does the seeing. Man or computer.? In the future a computer will see a tactical gain of a piece in 15 moves. If a person did the move because of strategic reasons then gained the piece, it will be a strategic move. The computer who saw the gain straight away it will be a tactical move.
Perhaps even today this occurs although less than 15 moves.
So does it depend upon the 'intent' of the player. ?
Taking ideas is not a vice, it is a virtue. We have another word for this. It is called learning.
But sharing ideas is an even greater virtue. We have another word for this. It is called teaching.
towforce wrote: ↑Wed Oct 30, 2019 5:11 pm
If one has built a fast-searching engine, one jolly well SHOULD tune it to seek tactical complexity!
I don't think so. Or, at least, that experiment has been tried. Tuning a chess engine to solve tactical test sets definitely lowers the strength in game play.
That's not quite what I meant. I intended to imply tuning the evaluation to prefer positions where lots of pieces are attacking lots of other pieces: the kind of position where there's lots of calculations to do each go, and little alternative but to sit and do the calculations. Such positions used to throttle off the search depth on older, slower computers, and I'd expect fast-searching engines to have a natural advantage in this type of position.
Writing is the antidote to confusion.
It's not "how smart you are", it's "how are you smart".
Your brain doesn't work the way you want, so train it!
towforce wrote: ↑Wed Oct 30, 2019 5:11 pm
If one has built a fast-searching engine, one jolly well SHOULD tune it to seek tactical complexity!
I don't think so. Or, at least, that experiment has been tried. Tuning a chess engine to solve tactical test sets definitely lowers the strength in game play.
That's not quite what I meant. I intended to imply tuning the evaluation to prefer positions where lots of pieces are attacking lots of other pieces: the kind of position where there's lots of calculations to do each go, and little alternative but to sit and do the calculations. Such positions used to throttle off the search depth on older, slower computers, and I'd expect fast-searching engines to have a natural advantage in this type of position.
I guess that almost all programs already do this. They give a bonus for mobility (and space, which increases mobility) and attacks advantage will be revealed already by the SEE function.
Taking ideas is not a vice, it is a virtue. We have another word for this. It is called learning.
But sharing ideas is an even greater virtue. We have another word for this. It is called teaching.
Ovyron wrote: ↑Thu Oct 31, 2019 8:25 am
A strategic move does not need calculation, you could tell them apart on a program if it finds the move without any search.
Two quotes for my archive on one page. This is an EXCELLENT thread!
Writing is the antidote to confusion.
It's not "how smart you are", it's "how are you smart".
Your brain doesn't work the way you want, so train it!
towforce wrote: ↑Wed Oct 30, 2019 5:11 pm
If one has built a fast-searching engine, one jolly well SHOULD tune it to seek tactical complexity!
I don't think so. Or, at least, that experiment has been tried. Tuning a chess engine to solve tactical test sets definitely lowers the strength in game play.
That's not quite what I meant. I intended to imply tuning the evaluation to prefer positions where lots of pieces are attacking lots of other pieces: the kind of position where there's lots of calculations to do each go, and little alternative but to sit and do the calculations. Such positions used to throttle off the search depth on older, slower computers, and I'd expect fast-searching engines to have a natural advantage in this type of position.
I guess that almost all programs already do this. They give a bonus for mobility (and space, which increases mobility) and attacks advantage will be revealed already by the SEE function.
The point is to prefer equal position when you need to calculate relative to equal position when you do not need to calculate.
programs with symmetric evaluation certainly do not do it and there are no positions that they prefer both for white and for black.
Uri Blass wrote: ↑Thu Oct 31, 2019 10:51 amprograms with symmetric evaluation certainly do not do it and there are no positions that they prefer both for white and for black.
Yes, Stockfish Contempt has shown asymmetric evaluation of positions is possible, and with high Contempt it can prefer the white side or the black side depending on who's it to move. You can build a tree where white has backsolved scores from leaf nodes of white contempt, and black has backsolved scores from leaf nodes of black contempt, allowing you to hold 4 PVs instead of 1 (1. Best white attack vs. Best black attack. 2. Best white attack vs. Best black defense. 3. Best White defense vs. Best black attack. And 4. Best white defense vs. Best black defense.) The best PV depends on the position, or your objective (you want the best defense if you are losing, it's useless to know your best attack if you know you can't win.)
Unfortunately this new paradigm has been impossible for some people to grasp, and they just turn off contempt to hold a single PV of symmetric evaluation and go on with their lives.
The fact that contempt makes engines play better in some positions is nothing more than proof that engines are far too passive when the current score is drawn.
Adding points to the existing score is a stupid way to solve it.
Taking ideas is not a vice, it is a virtue. We have another word for this. It is called learning.
But sharing ideas is an even greater virtue. We have another word for this. It is called teaching.
Draw value - This was how old engines used to measure "Contempt", all they did was changing the value of a draw from 0.00 to anything, so if you wanted a draw, you could set it to 2.00 and the engine would be willing to give away a 1.99 advantage if it could see a way to secure a draw. The obvious problem here is that values like -2.00 would confuse the engine and make it think that the opponent wanted a draw when it wasn't true, so something like this needs careful implementation.
Piece values - Back when Rybka was #1 engine, it came out with settings that would allow users to create personalities for it by making it change the value of the pieces, bishop pair bonus, rook endgame penalties and so on. One powerful aspect of it was allowing the engine to change the piece's values of each color individually (unlike others where a Knight value applied to both sides.) Nobody was able to come up with anything useful for those, except for Mindbreaker, who made one that would make Rybka's pieces more valuable than the opponent's pieces, and this worked wonders to make it play stronger against weaker opposition (which was everyone else, back then.) But it also allowed Draw value to be tweaked, so one could change it and make Rybka play moves improved by the settings but make it show an eval of 0.00 for draws and go for them accordingly. This is really poor against stronger opposition.
Contempt - What engines are doing now. The concept is the same, but engines have different implementations. In some positions high Contempt will make Stockfish keep material on the board while Komodo will want to exchange it down, for instance. Houdini's Contempt was dynamic and it was only applied to a side with the advantage. It's a dynamic way to decrease draws with it on, and it's sometimes the only way to find something the fastest in quiet positions.
I hold that an engine can have these three factors tweaked to make it go from the limits of tactical ability to strategical limits. What we haven't seen before is an engine that looks at a position and adjusts its own Draw Values, Piece Values and Contempt values to perform the best in it.
Because some positions require tactics, and others require strategy, perhaps something like Ferdy's Aiquiri could have the solution, by having a strategical engine and a tactical engine suggest a move, and have a third judge engine that knows what the position needs pick one of the moves.
