Author: Brian Richardson
Date: 05:54:51 08/27/00
Go up one level in this thread
On August 27, 2000 at 02:01:37, Peter McKenzie wrote: >On August 26, 2000 at 23:03:49, Will Singleton wrote: > >>On August 26, 2000 at 16:12:45, Peter McKenzie wrote: >> >>>On August 26, 2000 at 12:44:46, Will Singleton wrote: >>> >>>>On August 26, 2000 at 08:06:00, Brian Richardson wrote: >>>> >>>>>I think the information that many programs report, such as depth, score, PV, >>>>>etc) is helpful to see, and appreciate most programs doing so (particuarly when >>>>>their operators are online, or even some that have lists of opponents to >>>>>automatically kibitz). >>>>> >>>>>However, each program reports in a different manner. I would like to propose >>>>>_some_ reporting uniformity. I am not suggesting that standardizing _all_ >>>>>information and formats should be attempted. >>>>> >>>>>In particular, the search "depth" seems like a good place to start. Depth can >>>>>mean several different things. I would like to propose a depth reporting format >>>>>as follows: >>>>> >>>>>ply x(y/z) where x is the last full width (normal search) ply _completed_, >>>>>z is the deepest with extensions, and z is the absolute deepest ply reached >>>>>(typically in q-search). >>>>> >>>> >>>>I assume you mean y = deepest with extensions. >>>> >>>>I don't know about standardizing. I sort of like to compare the different >>>>formats, you get to know the idiosyncrasies of each after awhile. Your format >>>>is unique I think, kind of verbose. Others give a single number, like d=8, >>>>which is too sparse. I like my method (surprise :)), that gives the depth >>>>reached plus number of ply 1 moves examined at that depth. That shows exactly >>>>where in the ply the search terminated. >>> >>>I prefer to show DEPTH COMPLETED, plus the number of root moves searched at the >>>next depth. I think that depth completed is a pretty universal thing. Alot of >>>programs don't show how much has been done at the next search depth, partly >>>because this can be complicated by re-searches etc. >>> >>>In general, I think it would be great to agree on a standard format and a >>>standard set of information to be presented. >>> >> >>Sure, if possible. But it may be that different folks see what's important in >>different ways. >> >>So, when you display 7.10 depth, that means you've searched seven full ply, plus >>10 root moves of the eighth ply? > >Correct > >> When I display 7/10 depth, it means I've >>searched six full ply, plus 10 root moves of the seventh ply. That is, I >>terminated the search at ply 7, move 10. That seems like the proper way to do >>it. > >I was under the impression that the convention was the opposite, and that people >were most interested in completed ply. > >> >>Will I am most interested in completed ply too, but "completed" or the "insurance depth" mentioned in a related post will vary _a lot_ based on null move and other pruning techniques. Nonetheless, I would rather have some information, other than just that the search stopped somewhere in ply n. As for the n.m reporting, where n is the full ply completed and m is the number of moves at the next ply searched, that is also useful infomation. However, it does not say how many more moves there are--so how about making m the number of root moves left to be searched. This tells us how close the search was to making the next ply (as in Winboard "analyze" mode). Of course that last move to be searched could turn out to be the best, and the time per move would not actually be even, but it is some additional information. Do others feel it is valuable? When Tinker is playing on ICC, I like seeing the kib's of other programs. I would rather see more depth and/or moves searched reporting than an excessively long PV, if verbosity is an issue. One awkward thing for me right now relates to predicted moves. Generally, if Tinker ponders for longer than the normal next move search time and the predicted move is made, then Tinker moves immediately. Unfortunately, the score and PV from the pondering are lost and bogus values are reported from the abnormally short search. This is annoying when reporting, and a real problem when going back to do position learning. I am thinking of just saving the pondering search results separately for these cases. Any other suggestions? The situation is similar but worse when there is only _one_ legal move, which Tinker makes immediately. I was reporting the raw root eval(), but this caused Tinker to resign a KNNvK game (which Tinker had seen was drawn according to EGTB's). I suppose EGTBs should be checked too. Finally, a different format would apply to book and/or learning moves, of course. I'll raise the issue even though Tinker _still_ has no book. Thanks, Brian
This page took 0 seconds to execute
Last modified: Thu, 15 Apr 21 08:11:13 -0700
Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.