Computer Chess Club Archives




Subject: Re: Couple of chess programming questions

Author: martin fierz

Date: 13:06:51 09/10/02

Go up one level in this thread

On September 10, 2002 at 15:41:42, Vincent Diepeveen wrote:

>On September 10, 2002 at 15:19:21, martin fierz wrote:
>>On September 10, 2002 at 14:45:27, Omid David wrote:
>>>On September 10, 2002 at 14:30:56, martin fierz wrote:
>>>>On September 10, 2002 at 09:26:14, Eli Liang wrote:
>>>>>A couple of chess programming questions:
>>>>hmm, i only wrote a checkers program, but here's my take:
>>>>>(1) Are there any uses for ProbCut and/or Multi-ProbCut in chess positions where
>>>>>the variance of leaf-nodes is low?
>>>>i've tried multi-probcut and it works well in checkers. i never tuned it as much
>>>>as my own pruning algorithm, and it doesn't perform quite as well - but it is BY
>>>>FAR better than no pruning. i'll be trying to tune it in the near future. for
>>>>games where the eval doesnt swing wildly, MPC is a fantastic algorithm.
>>>>>(3) Reading Aske Plaat's search & re-search paper, it really seems like mtd(f)
>>>>>is something of a magic bullet.  But I note it seems that more programs don't
>>>>>use it than do (for example Crafty).  What is wrong with mtd(f) which Plaat
>>>>>doesn't say?
>>>>i'm using MTD. i tried windowed search, PVS and MTD. in my tests, in long engine
>>>>matches, MTD performed marginally (no statistical significance...) better than
>>>>PVS. it typically searched a low 1-digit % less nodes for a given depth than
>>>>i don't know how to get a PV out of MTD. in normal searches, a pv node is where
>>>>the value is > alpha but < beta. in MTD, you never get this condition.
>>>>retrieving a PV from the hashtable is possible, but in all probability, you will
>>>>not get the full PV. which is real bad for debugging if you want to know what
>>>>the program was thinking at the time... i once asked here how to get a pv from
>>>>MTD but got no answer - and if you can't get the pv, then that is a major
>>>I haven't tried getting the PV out of MTD(f), but just a thought: why should
>>>there be any problem in getting the PV out of hash table? Play the first move,
>>>update the position, get the next best move from hash table, and so on... ?!
>>there's no problem with that except that on any reasonably deep search, you will
>>not have been able to store all pv nodes in the hashtable. so you end up with a
>>search which says it was 23 ply deep and have e.g. 15 pv moves. if you just want
>>to display it for the user, that's fine. but if your program plays a bad move,
>but then your hashtable management sucks ass, sorry to say so.

but you don't use MTD! which means you *know* when you have a pv node, because
"pvnode <=> alpha<value<beta". and then you can make sure it doesn't get
overwritten in the hashtable. if you use MTD, you don't have this information -
all your hashtable entries are either lower or upper bounds... so how do i know
which ones i have to keep? i'd really glad to learn how to do this :-)
so if you can tell me how to do it instead of saying i suck (well possible...),
i'd love to try!


>I get in Napoleon also only mainlines out of hashtable (with pvs)
>wasting system time in the search to update all kind of stupid
>arrays for it is a waste of time, and the next iteration you get
>true bounds, so you can't get the mainline in arrays anyway (mtd
>is different here). finding a win in 50 ply is no problem to display...
>>and you want to know what line it was considering as being best, e.g. because
>>you want to know if your static eval is bad in the final node of the pv, you
>>can't do it. IMO debugging your program and finding eval problems like this is
>>MUCH more important than something like 5% more speed.
>>  martin
>>>>>(6) Has anyone found any real "practical" benefits to fractional-ply extensions?
>>>>yes. i tried recapture extensions of different depth, and half a ply gave the
>>>>best result. don't ask me why, it's just an observation.
>>>>  martin

This page took 0.03 seconds to execute

Last modified: Thu, 07 Jul 11 08:48:38 -0700

Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.