Blunder related to aspiration window?

[sandermvdb]

Sometimes my engine makes a huge blunder and I was wondering if this is related to the aspiration window. Suppose I have no time left and MUST perform a move, what should I do in the following situations? Perform the best move of the previous iteration or perform the 'best move' of the current iteration?
1) root score <= alpha
2) root score >= beta

[syzygy]

If the search has definitely proved that, at the current iteration, the current best move is better than the best move of the previous iteration, then play the current best move.

In your case the search may only have detected a null-window search that failed high. It may be better to first verify the initial fail high with a PV search before promoting the move to new best move.

[mar]

Sometimes my engine makes a huge blunder and I was wondering if this is related to the aspiration window. Suppose I have no time left and MUST perform a move, what should I do in the following situations? Perform the best move of the previous iteration or perform the 'best move' of the current iteration?
1) root score <= alpha
2) root score >= beta

What I do, if aspiration fails low, give the engine more time (certain fraction of remaining time) to resolve the iteration (this is typically a nice elo gain as well). (if time's up even after this then I play the previous best move)
A subtle detail is that when failing low I resolve the failing move first rather than looking for a better one with current bounds.
When failing high, I happily play the move betting that it will be good (in most cases it is). Some resolve the fail high as well.

[sandermvdb]

What I do, if aspiration fails low, give the engine more time (certain fraction of remaining time) to resolve the iteration (this is typically a nice elo gain as well). (if time's up even after this then I play the previous best move)
A subtle detail is that when failing low I resolve the failing move first rather than looking for a better one with current bounds.
When failing high, I happily play the move betting that it will be good (in most cases it is). Some resolve the fail high as well.

What do you mean by 'resolve the failing move'?

And suppose I do play the failing move, as I do now, could this sometimes result in huge blunders? Giving away the queen for instance.

[Joost Buijs]

Sometimes my engine makes a huge blunder and I was wondering if this is related to the aspiration window. Suppose I have no time left and MUST perform a move, what should I do in the following situations? Perform the best move of the previous iteration or perform the 'best move' of the current iteration?
1) root score <= alpha
2) root score >= beta

What I do, if aspiration fails low, give the engine more time (certain fraction of remaining time) to resolve the iteration (this is typically a nice elo gain as well). (if time's up even after this then I play the previous best move)
A subtle detail is that when failing low I resolve the failing move first rather than looking for a better one with current bounds.
When failing high, I happily play the move betting that it will be good (in most cases it is). Some resolve the fail high as well.

Basically I do the same, when the PV move fails low I give the engine additional time and resolve the move with adjusted bounds before continuing with the non PV moves. When a non PV move fails high, I usually resolve it, only with very little time left on the clock I play the FH move immediately.

[Joost Buijs]

What I do, if aspiration fails low, give the engine more time (certain fraction of remaining time) to resolve the iteration (this is typically a nice elo gain as well). (if time's up even after this then I play the previous best move)
A subtle detail is that when failing low I resolve the failing move first rather than looking for a better one with current bounds.
When failing high, I happily play the move betting that it will be good (in most cases it is). Some resolve the fail high as well.

What do you mean by 'resolve the failing move'?

And suppose I do play the failing move, as I do now, could this sometimes result in huge blunders? Giving away the queen for instance.

I don't think that when the PV move (the best move from previous iteration) fails low it should play such gross blunders, it must be some other kind of bug. Maybe there is something wrong with your move sorting.

[sandermvdb]

I don't think that when the PV move (the best move from previous iteration) fails low it should play such gross blunders, it must be some other kind of bug. Maybe there is something wrong with your move sorting.

I am using 2 transposition tables and using the score/move with the highest depth. Maybe that is the cause.

[mar]

What do you mean by 'resolve the failing move'?

I mean resolving the iteration if time permits by using more time.
If you can find a better move than the failing move that resolves within aspiration bounds then you play it,
that's why I resolve the failing move first by adjusting bounds to see how bad it is.
EDIT: by failing I mean failing low

[Joost Buijs]

I am using 2 transposition tables and using the score/move with the highest depth. Maybe that is the cause.

So I assume you get the best-move and ponder-move from the transposition table. In the past I also did this and occasionally it returned the wrong move, since that time I treat them separately.

[JVMerlino]

Sometimes my engine makes a huge blunder and I was wondering if this is related to the aspiration window. Suppose I have no time left and MUST perform a move, what should I do in the following situations? Perform the best move of the previous iteration or perform the 'best move' of the current iteration?
1) root score <= alpha
2) root score >= beta

There is no one right answer, because this is actually a time management question. Yes, you must eventually either perform the best move of the previous iteration or the best move of the current iteration. But you don't have a best move of the current iteration until the first move you've searched in that iteration comes back with a true score (as opposed to failing high/low and needing to be re-searched).

So, the question becomes how to best use your time. The wrong solution would be to always use the same amount of time per move (typically 1/N of the clock remaining at the start of the search - usually N is somewhere between 30 and 40 in rudimentary implementations).

Myrddin is very aggressive with its clock use. If the first move of the current iteration fails low, I will use up to half of the remaining clock in the hopes that it can be resolved. But this calculation is also potentially affected/adjusted by the current score, and by the relation of the current score and the score of the previous iteration. For example, if Myrddin clearly thinks it's winning, it certainly won't spend a bunch of extra time trying to get an exact score.

I'm pretty pleased with the results. Myrddin almost always uses more of its clock than its opponents. Sometimes it will get itself into time trouble, but isn't it better to resolve those fail lows when in crucial positions?

But, as to your initial concern about playing blunders, aspiration windows shouldn't cause that. I can only guess that, if you run out of time and haven't yet found an exact score for any move in the current iteration, then perhaps your intention is to play the first move searched, but are playing some other move instead due to a bug?

jm
[/b]