Author: Amir Ban
Date: 15:53:19 11/25/99
Go up one level in this thread
On November 24, 1999 at 00:36:41, Michael Neish wrote: > >Hi, > >I started this thread, "The Limits of Positional Knowledge" about a couple of >weeks ago. I'm happy to see that it's still ticking over, although by now it's >evolved pretty far from what my original question was! > >A couple of the early replies to my question were on the mark, but >the rest have strayed into areas far more noble and sophisticated than my >lowly inquiry. But since no one's really answered my (naive) question -- >maybe there is no answer? -- I'll restate it. > >Having recently written my first program, which can search only up to 6-ply, >that's if you want it to move sometime before the big crunch, I thought what >would happen if I stubbornly kept to a 6-ply limit and just tried to improve >the program's playing strength by giving it increasingly better positional >sense? >I presume that given perfect positional knowledge no lookahead at all would be >required, since all tactical opportunities would be nipped at the bud (although >how it would exploit tactical blunders by the opponent is a separate question in >itself). So my question was "how strong can a program become just by >concentrating on the positional evaluation?". I will allow the use of >extensions >if a tactical possibility is discovered. > >I fiddled with my program with this in mind, and although I found that it seems >to ponder over better moves quicker with a better evaluation function, I seem >to be hitting a wall as far as performance improvement is concerned. For >instance, if I play the program against itself, with White adopting the old >parameters, and Black the new, it seems Black plays better, but after developing >an advantage it cannot convert. > >I appreciate any replies. Thanks. > >Mike Maybe the reason you didn't get real replies to your question is that you are asking a really good question. I identify with you, because I've been asking myself the same question during the last year. Most of the effort that went into Junior 6 was spent on systematic improvement of the program's evaluation. I spent on this much more time than I spent, say, on the Deep Junior (SMP) code. While doing this effort I often wondered whether this is really an effective way of taking the program one level higher, and at other times I wondered what will be the limit of this approach. I'm not sure what the answers are, but you will be able to see the effect in Junior 6 as compared to Junior 5, though I did spoil the clean comparison somewhat by changing other things. As a more direct answer to your question, if you have a 6-ply searcher with an excellent and superior evaluation, you will have a nice program. Whether you will win will depend very much on how much your better evaluation will show, or to put it differently, how many positional errors your opponent will make, and how severe they will be. It's well known that with a sufficient advantage you can win even against huge tactical power, because the game is, well, won. When you have a big advantage the deep searcher will often only see the end sooner. The problem is that you have to get a big advantage in the first place, and this must come as a gift from the opponent. If the game is near-level and you are tactically weak, you may get killed tactically before the opponent makes a positional error of any significance. The problem with today's top programs is that they don't make enough such errors to offset more than slightly tactical inferiority (though they do make such errors). Here's something I heard Garry Kasparov say: If a strong human is allowed to override just one move in a game I play against a computer, I'm already in big trouble. Since Garry's positional knowledge is probably the best there is, you have in this sentence the opinion of the world champion on the limits of positional knowledge. Another point: If you insist on having a 6-ply search, better make it a well-tuned 6-ply search with good extension strategy and good quiescence and control of the horizon effect. If you don't have that, even the best evaluation won't save you from a quick loss. Amir
This page took 0 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.