Moderators: hgm, Rebel, chrisw
Code: Select all
do {
score = search(..., depth, alpha, beta)
if (score <= alpha)
alpha=-INF;
else if (score >= beta)
beta = +INF;
} while (score <= alpha || score >= beta);
// set aspiration window for the next depth
alpha = score - 150;
beta = score + 150;
According to my experience aspiration windows will add some elo when done properly.lucasart wrote:I've been doing quite a few experiments with those, and could never come up with anything elo increasing. So I currently do a search with (alpha,beta) = (-INF,+INF).
I wonder how other engine developers do that, and if they managed to find a way that actually improves anything.
Amongst my several attempts, the most "logical" one would bethis pseudo code is of course embedded in the iterative deepening loop.[/code]Code: Select all
do { score = search(..., depth, alpha, beta) if (score <= alpha) alpha=-INF; else if (score >= beta) beta = +INF; } while (score <= alpha || score >= beta); // set aspiration window for the next depth alpha = score - 150; beta = score + 150;
It looks it will jump out of the loop after the first iteration. It may also be better to do alpha=max(-INF, alpha-margin); beta=min(+INF, beta+margin). I reset to -INF,+INF immediately after a fl or fh because, until now, I could not test an improvement by gradually opening the window and code is (a tiny bit) more readable with simple method to open it fully at once.lucasart wrote:I've been doing quite a few experiments with those, and could never come up with anything elo increasing. So I currently do a search with (alpha,beta) = (-INF,+INF).
I wonder how other engine developers do that, and if they managed to find a way that actually improves anything.
Amongst my several attempts, the most "logical" one would bethis pseudo code is of course embedded in the iterative deepening loop.[/code]Code: Select all
do { score = search(..., depth, alpha, beta) if (score <= alpha) alpha=-INF; else if (score >= beta) beta = +INF; } while (score <= alpha || score >= beta); // set aspiration window for the next depth alpha = score - 150; beta = score + 150;
Yep.zamar wrote:AFAIK the recipe that all top engines are using is:
- Use very small aspiration window (only something like 8cp - 16cp).
- Use very low effective branching factor (~1.6)
Yep.This will result in many researches, but for some reason it seems to pay off. Also you must note that with low EBF + small aspiration window, you may want to adjust your time management to be more aggressive compared to high EBF without aspiration window.
IMO all these three issues (aspiration window, EBF and time management) are so closely related that you cannot tune one without also tuning the others.
Sure! Old chess programs use various extensions and few reductions. This results in high-branching factor >= 2.0.jacobbl wrote:You are saying one should use a low ebf, just as if it's a parameter you can choose. One of my largest struggles is to lower my ebf. Could you please enlighten me on what you mean with: "Use a low ebf"?
Regards
Jacob
In the pseudo code given above the "search(...)" call is meant as the search of the root node, it is the code for one single iteration, i.e. (part of) the body of the iteration loop:Evert wrote:Correct me if I'm wrong, but you only need an exact score for the best move, so for any other move you don't adjust the window and research when they fail low because that's the expected behaviour. Right?
Your pseudo-code sample seems to always research when a move fails low, even if you expect it to.
Or am I completely missing something?
Code: Select all
for (iteration = 1; iteration <= maxIteration; iteration++) {
// now the code above
}