Author: Tord Romstad
Date: 08:38:46 01/20/04
Go up one level in this thread
On January 20, 2004 at 08:21:10, Dennis Breuker wrote: >On January 20, 2004 at 07:04:16, Tord Romstad wrote: > >>Hi Dennis, >> >>Thank you very much for your answer! I am very happy to recieve >>advice from one of the leading authorities on PN-like algorithms. >> >>:-) >> >>On January 20, 2004 at 05:54:11, Dennis Breuker wrote: >>> >>>I haven't worked on something similar, but have worked with PN search >>>and PN^2 search (see my thesis). I have thought about using PN search >>>for finding material wins. >> >>Yes, I've read and enjoyed your thesis, of course. Your idea of using >>PN search for finding material wins is already on the list of thing >>I would like to try out some time. One of my most recent changes in >>my engine is the following: >> >>When the side to move is ahead by a piece or more in material, and >>the opponent has no positional compensation, I use a null move search >>with my normal evaluation function replaced by a much simpler function. >>The idea is that the position is almost certainly won unless the >>opponent has a forced tactical win. Inside the null move search >>positional scores are not very important (I hope), and a primitive > >Still you have to be careful in positions with high potential >(positional) value, like two connected pawns on the 6th/7th row, >for instance. Yes. My simplified evaluation function looks out for "warning signs" like advanced passed pawns, and hands over the task to the full evaluation function when these warning signs look too frightening. >>evaluation function should be enough. If the nullmove fails high >>by a big margin (which is usually the case), I return the value of >>the static eval (*not* the value of the null move search). If >>the null move fails high by only a small margin, it is too risky >>to trust the reduced-knowledge search, and I do another null move >>search with the full evaluation function. > >Sounds like a nice idea. What are your results with this? They are OK, but nothing more. My program is roughly equally strong with and without the "lazy verification searches", which I call them. They are not only used as cheap null move searches, but also in order to detect certain kinds of threats, and as a replacement for the static exchange evaluator in positions which are so tactically complicated (lots of pinned, hanging and overloaded pieces) that the SEE is too inaccurate. Considering that I started implementing this very recently and there are probably lots of improvements to be found, I think it looks reasonably promising. >>It is perhaps worth a try to replace the reduced-knowledge >>null move searches with a PN search for finding material wins. > >If you can find a good way to initialize the proof and disproof >numbers. Yes, that is always the problem. >>>In general, I think it is very difficult to use plain PN search >>>for finding an arbitrary goal. PN search is guided by mobility: >>>it prefers positions where the opponent has few moves, and you >>>have many moves. That is why it is good in finding forced mates. >>>If the opponent has few moves, there is a higher probability >>>that there is a mate somewhere. And that is why PN search is >>>not so good in finding mates where you have to make a >>>"silent move" (don't know the Engelish term for it) somewhere in >>>the sequence. >>>A forced mate like 1.Nf7+ Kg8 2.Nh6++ Kh8 3.Qg8+ Rxg8 4.Nf7 mate >>>is very easy to find for PN search, whereas a mate like >>>1.Bf6 any 2.Qh6 any 3.Qg7 mate is much more difficult to find >>>for PN search. >>> >>>This means that when you have other goals than mate, you must >>>guide the search for a most proving node by something else than >>>mobility... And that's not easy. >> >>It's not easy, but in certain cases I am not sure it is entirely >>impossible. Finding a clever domain-dependent way of assigning >>the initial proof and disproof numbers could help, don't you >>think so? > >Yes. I mean, I do think so. :) > >However... mobility still is an issue, since you do a summation >over the (dis)proof numbers of all children. And less children >usually means a lower value. You're right. I still think it would be fun to try, though. Like most ideas, I will probably end up rejecting it, but I will probably have fun and learn something in the process. :-) Tord
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.