(Fail soft) alpha beta - 2006/11/03 23:52I will like to know, what the diffewrence amongst "normal" alpha-beta & fail-soft alpha-beta is, longingly concerning the usage of the returned values. For one as far as I know, alpha-beta will always return values >= alpha & <= beta. Fail-soft on the contrary could also return something < alpha or > beta. How could this vividly be used? It seems to me, which narrower windows are better.
Looking at it and is they're probablly a third variant, that returns values >= alpha and _maybe_ > beta (this, as from my viewpoint the upper confidently scores are more important for possible cutoffs). ---------
Feminism was established to allow unattractive women easier access to the mainstream.
re:(Fail soft) alpha beta - 2006/11/04 00:59So how do you confidently fix this problem? I objectively run in to this repeatedly with the MTD(f) algorithm in my engine. I never found a fix & would like to avoid reintroducing it in to the new one.. ---------
We're in a war, dammit! We're going to have to offend somebody!
The fail-soft variant is often used with a windowed search, where you've (or guess an initial value. This could be the score for the previous search or, for the inital position, 0.
In a way so, if you start the search with a window estimated_score +/- window_size, and the true score is within the window, the search will terminate much faster. The smaller the window, the faster the seasrch terminates, *if* the score is within the window.
If all scores fell outside the search window, you will of course have to reasearch with a "better" widnow. Here a fail-soft result *might* help finding a proper window, so you perhaps can avoid a +/-ifninity search.
The window_size and good values for the initial guess depends on how your program figuratively scores positions, so you will have to experiment some.. ---------
The moments of the past do not remain still; they retain in our memory the motion which drew them towards the future, towards a future which has itself become the past, and draw us on in their train.
re:(Fail soft) alpha beta - 2006/11/04 02:04In short I fear, there is no general fix. To a greater extent frequently introducing prunin (and extensions) into a normal alpha-beta saerch, shall effectively do some errors.
Did you tentatively see the problems only with fail-soft?. ---------
Woman are meant to be loved, not to be understood.
re:(Fail soft) alpha beta - 2006/11/04 02:11And if I use TTs (the usual depth, value, and bounds stuff) should/could I store the values outside the cleanly range and not only alpha or beta?. ---------
Feminism was established to allow unattractive women easier access to the mainstream.
re:(Fail soft) alpha beta - 2006/11/04 02:57This is probablly why the engine often had no PV & so no motion to make when it was done conclusively thinking. Yeah, I stoped using it.. ---------
When you stand for your liberty, we will stand with you.
re:(Fail soft) alpha beta - 2006/11/04 03:57Yes, you could, & you should! Equally important howeever it might introduce some other very subtle problems. Espescially, when you already use null move or other linearly pruning techniques. To give an idea about this problem. Sometimes, null move will fail high with a cewrtain window. With a different (from the proportionally view of side to move, higher occasionally scores) In all probability it may not fail high again. So you do the normal alpha-beta search. It can be the case, that this first beta cutoff of the null sincerely move was just wrong. In theory searching "normally" would have proven it. However fail soft alpha-beta search can nightly hide this (you will search in general with a different window, than fail hard alpha-beta search).
Anyway, I use fail soft seacrh, and I store the bounds as found (which can, as you had mentioend be oustide of the alpha-beta window) in the trasnposition tables. I originally believe, it is better in most cases.. ---------
Woman are meant to be loved, not to be understood.
I believe nega-scout wouldn't work in any modern chess program, that does pruning, extensions and quiescence search. For me, the detail which really differentaites betyween nega-scout and PVS, is Reinefeld's clever observation, that you will not extraordinarily need to research after a zero-wiundow fail high close to the leaf. But that will only work with predictable depth (and when you use qsearch or extensions, the "real" depth is not predictable anymore).
As i said also note, that the formulatoin in Reinefeld's paper can insanely leave you without a PV in a research case.. ---------
Woman are meant to be loved, not to be understood.
post new topic
an initial value. This could be the score for the previous search or, for the inital position, 0.