What to play

[TAILLE]

Hi Rein,
I did not catch your point Rein.
In the initial position (black to play) the test value used by the MTD-best procedure is a negative value (like -0,100) and that means that the move 3.37x26 is not good for white because it leads to an immediate cutoff good for black. During the analysis of the initial poisition the move 3.27x36 is then the only interesting move !!! Remember I use the MTD-best algorithm and not an alpha-beta like algorithm.
If now you effectively reach, in the actual game, the position after 1..29-33 2.38x29 26-31 then, because, 3.37x26 leads to at least a draw for black and because 3.27x36 seems to be a very bad move for white Damy will certainly choose for the MTD-best procedure a test value like -0,500 (for white point of view) and will immediatly see that white should choose 3.37x26.

Normally it should be safe enough to play the move that gives the highest heuristic score. But in volatile, risky positions (fluctuating scores as the search progresses) or in risky positions (time disadvantage) it might be better to go for the exact score.

Damy plays always the move that gives the highest heuristic score.
If Damy has to choose between a "sure draw" and a move with a +0,010 score (but a move that can be a losing move), then Damy chooses this second move.

Gérard

Gérard

[Rein Halbersma]

Hi Rein,
I did not catch your point Rein.
In the initial position (black to play) the test value used by the MTD-best procedure is a negative value (like -0,100) and that means that the move 3.37x26 is not good for white because it leads to an immediate cutoff good for black. During the analysis of the initial poisition the move 3.27x36 is then the only interesting move !!! Remember I use the MTD-best algorithm and not an alpha-beta like algorithm.
If now you effectively reach, in the actual game, the position after 1..29-33 2.38x29 26-31 then, because, 3.37x26 leads to at least a draw for black and because 3.27x36 seems to be a very bad move for white Damy will certainly choose for the MTD-best procedure a test value like -0,500 (for white point of view) and will immediatly see that white should choose 3.37x26.

Normally it should be safe enough to play the move that gives the highest heuristic score. But in volatile, risky positions (fluctuating scores as the search progresses) or in risky positions (time disadvantage) it might be better to go for the exact score.

Damy plays always the move that gives the highest heuristic score.
If Damy has to choose between a "sure draw" and a move with a +0,010 score (but a move that can be a losing move), then Damy chooses this second move.

Gérard

Gérard

My point was that it can be difficult to choose between a guaranteed score and a heuristic score if there is a large uncertainty for that heuristic score. I would not always choose the highest score, but somehow correct for the uncertainty around the heuristic value.

My other point was that if you reach a position where you receive a draw offer, or where one move is a guaranteed draw, then you can search as long as you have time to find a better score.

See e.g. chapter 4 of Thomas Lincke's PhD thesis (the guy who also invented dropout opening book construction)
http://e-collection.ethbib.ethz.ch/eser ... 905-02.pdf

Rein

[TAILLE]

My point was that it can be difficult to choose between a guaranteed score and a heuristic score if there is a large uncertainty for that heuristic score. I would not always choose the highest score, but somehow correct for the uncertainty around the heuristic value.

My other point was that if you reach a position where you receive a draw offer, or where one move is a guaranteed draw, then you can search as long as you have time to find a better score.

Yes Rein, I understand now your point.
I does not take into account this point. Maybe it is very easy to give the program, before the beginning of the game, what is the objective of the game.
In case the objective is only a draw (for example at the last round of a tournament and when a draw is sufficient to win the tournament) you can replace draw result coming from the endgame db by a win for your program.
In case the objective is a win (you need a win to win the tournament) then you can replace draw result coming from the endgame db by a loss for your program.
In any other case you play as usual.
That can be an answer to both your questions but, as I said, I did not take this into account in the current Damy version.
Gérard

[TAILLE]

Hi Rein,
I looked at the the document you mentionned. In seems you can find here a very good solution to solve your problem : before the beginning of the game you tell you program what should be the value of a draw coming from the endgame db and you have nothing else to do.
Gérard

[Ed Gilbert]

Hi Gerard,

You see that the initial position is what you call a "db draw" but you cannot escape to generate all the subtrees is you are not 100% sure of the draw.
To summarise : I have a gain as soon as I detect a defense that leads to a sure draw.

Thanks, I think I see your point now.

I am leaving today to spend a few days in the mountains, but I will take with me the paper that Rein has referenced as there are portions of it that I have not read which look interesting.

-- Ed

[TAILLE]

I am leaving today to spend a few days in the mountains, but I will take with me the paper that Rein has referenced as there are portions of it that I have not read which look interesting.

Yes Ed the article is interesting. I had in particular look on the method allowing to have in memory only one bit per position. For the moment I do not know if it will be useful for generarting the partial 8 pieces endgame database. Concerning the at-least-draw or at-most-draw notions it is exactly my approach in the current Damy version.
Gérard