Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: Future of Chess: Will GMs be able to draw computers?

Author: Vasik Rajlich

Date: 02:24:30 10/20/04

Go up one level in this thread


On October 20, 2004 at 05:08:13, Uri Blass wrote:

>On October 19, 2004 at 12:45:07, Vasik Rajlich wrote:
>
>>On October 19, 2004 at 09:34:00, Tord Romstad wrote:
>>
>>>On October 19, 2004 at 05:57:06, Vasik Rajlich wrote:
>>>
>>>>It's a sort of interesting accident that computer vs human is balanced at the
>>>>moment. If the game was more tactical, humans would already be crushed and we
>>>>would accept it as a matter of course. If it was more positional, humans would
>>>>still be stronger, maybe much stronger.
>>>
>>>I don't think it is quite that simple.  Consider the game of shogi, which is
>>>arguably a more tactical game than western chess.  The best computers have no
>>>chance against the best humans.
>>>
>>>I think there are two main reasons why computers are so good at chess:
>>>
>>>1. Material is very important in chess.  A material advantage is usually
>>>sufficient to win.  This makes it easy to write reasonably accurate evaluation
>>>functions for chess.  In shogi, material is much less important, while factors
>>>like initiative and king safety (which are much more difficult to quantify)
>>>play a more significant role.
>>
>>True - although this is something which helps make chess more tactical and less
>>strategic, as I understand those words.
>>
>>>
>>>2. The branching factor in chess is not too high.  This means that computer
>>>programs running on current hardware can afford to waste lots of time analysing
>>>nonsense variations, and still search deeply enough to see the relevant
>>>tactical lines.  This is not yet the case in shogi.  The horizon effect is
>>>still a very real program, and top human players are able to outsearch the
>>>best programs in tactical lines.
>>>
>>>Even in chess, I am not sure it is strictly true that computers are better
>>>than the best humans in tactics.  The picture is a bit more complicated.
>>>Computers are extremely good at finding short, bushy tactics, while the
>>>best GMs are better at finding very deep, narrow tactics including non-checks
>>>and quiet moves near the end of the lines.  Consider the Nolot test suite,
>>>which consists of problems solved by human players at a normal tournament
>>>time control, but which the computers still struggle to solve.
>>>
>>
>>Yes, it's probably an oversimplification to say that computers can calculate and
>>humans can evaluate. Still, in practice, in chess, it's more-or-less true. The
>>deep narrow tactics are not so common or so important. I don't see people losing
>>because of this, or computers losing to humans because of it.
>>
>>>In chess, computer hardware has reached the stage where the programs'
>>>superiority in short, bushy tactics almost exactly compensates for their
>>>relative weakness at understanding deep, narrow tactics.  In shogi, we
>>>will have to wait a few more years before this happens.
>>
>>Actually, there is one other property which chess has which helps make chess
>>engines strong. Given a primitive evaluation, as you search more, you play
>>better positional moves. I am not sure that this is inevitable for any game.
>>
>>There is a related issue, that of validating changes to evaluation. Evaluation
>>terms can be broken down into:
>>1) terms which are relatively dynamic, and can be viewed as approximations to a
>>deeper search (for example, king safety)
>>2) terms which are relatively static - and not related to search (for example,
>>pawn structure)
>
>I disagree here.
>
>If you know nothing about passed pawns you still can see often if you search
>deeper that passed pawns give you positional advantage in mobility.
>
>I do not think that king safety is something that can be replaced by deeper
>search better than pawn structure.
>
>Another point is that the question if some term is related to search is
>dependent on your other terms so you cannot divide them by that way.
>
>Uri

I really should have said "some aspects of pawn structure" rather than just pawn
structure.

For example, let's say that white has a protected passed pawn on a7, and the
evaluation doesn't give it any bonus. In the middlegame, all the search in the
world won't ever figure out that it should have any value.

Most chess evaluation terms are dynamic, though - ie. search can replace them.

Vas



This page took 0.01 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.