Author: Robert Hyatt
Date: 18:55:42 11/02/98
Go up one level in this thread
On November 02, 1998 at 19:04:26, Inmann Werner wrote: >On November 01, 1998 at 13:03:07, Inmann Werner wrote: > >>On October 31, 1998 at 19:53:46, Andrew Dados wrote: >> >>> >>>On October 31, 1998 at 16:45:01, Inmann Werner wrote: >>> >>>>I have problems with "nearly mate" positions, where the opponent can do some >>>>checks, so the mate gets beyond the horizon. >>>> >>>>One Example (really interesting!) >>>> >>>> >>>>Position (LCTII CMB 01) >>>> xrxxxxkx >>>> xqxxbpxx >>>> xxxpxxpx >>>> xxpPxxxx >>>> PxNxxxxR >>>> xxPxxQPx >>>> xPxxxPKx >>>> xxxxxxxx >>>> >>>>Best: c4d6/e7d6/f3f6 white wins >>>> >>>>In depth 7 , my program should find the solution, but it does not! The problem >>>>is, that black can make checks to bring the mate beyond the horizon. At each >>>>check, I extend the search for 1 ply, but it is not enough. >>>> >>>>The computer thinks: >>>>C4d6/e7d6/f3f6/b7d5+/g2h2/d6g3+!!/h2g3/d5d3+/g3h2/g8f8/ no h4h8 and black >>>>"stands good" >>>>7 plys plus 3 plys the checks end up at 10!! >>>> >>>>This things really often happen, and I am reall unhappy with it. When I tried to >>>>extend at a check with 2 plys, this problem was solved very early. But the >>>>overall searchdepth goes dramatically down, so the program plays worse. >>>>On an attack, it is nearly always possible for the opponent to slow down the >>>>attack with "senseless", but forced moves. >>>>I thought about something like: >>>>This is a mate in 1, therefore don´t let you disturb with "senseless " moves. >>>>But I have no thought, how to implement such a thing. >>>>Suggestions? >>>> >>>>Werner >>> >>> One solution I find acceptable is whenever you make null move generate all >>>checks at next ply even if null throws you to quiescence... that way null >>>returns 'being mated' score - and based on that you can extend your search >>>deeper...hopefully going beyond all delaying moves. >>> >>>Andrew >> >>I also thought about using the null move for this problems. Now my program finds >>the move at ply eight. >>I do not agree, that "brute force and fast" going into deeper plys is always >>better. If you make some "nice things", maybe the program finds something one >>ply earlier, which means "speed increasing" by factor three. >>So its worth to think about it, but always calculate the costs. >>I will try something with null moves and extensions, lets look what comes out. >> >>Werner > >********* >I tried something with null move. > >Problem: you can make a mate with the next move, but the opponent checks you as >long as it goes beyond the horizon. >I now only talk about this specific problem. I think, it often occurs and know, >that a "solution" for this problem does not stop horizon effect all. > >My thought: When the opponent is on move, I do a null move and one own move and >then quiet search. If a mate occurs here, I remember it, and pass it to the next >alpha_beta recursion which is my move. Only if a am checked at this time, I >extend 2 moves (not only one for the check) as long as extensions are allowed. >(Max Extensions...) >So I only extend if I can mate in one and the opponent slides out with a check! >There are also some limitations when doing the special null move, but there I am >experimating. > >My position above then gets solved at ply 6!!!!! >I tried some other positions with heavy threats and checks, where no real mate >can be made (worse for my algoritm). There it slows down about 10%, because it >looks deeper in maybe interesting things which in this case do not work. > >I must say, that I implemented above 3 hours ago, and 3 hours testing is not >much, but it is worth to think about. > >Werner my advice is "caution" doing this. IE Crafty finds this at depth=10, about 10 seconds or so on my ALR. It's not hard to make it find it at depth=6 or 7. But would it play better in games? No... it is *easy* to over-extend and find wonderful tactics but get killed positionally...
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.