Author: Albert Silver
Date: 06:47:40 02/12/00
Go up one level in this thread
On February 12, 2000 at 02:29:36, Dave Gomboc wrote:
>On February 12, 2000 at 02:00:09, Albert Silver wrote:
>
>>On February 12, 2000 at 01:38:02, Vincent Vega wrote:
>>
>>>On February 11, 2000 at 20:32:31, Albert Silver wrote:
>>>
>>>>It's still very complicated even where computers are concerned. I am presuming
>>>>this is a project of your own as you did not mention any names of third parties
>>>>doing the analysis, in which case I'd like to ask how these conclusions will be
>>>>reached.
>>>
>>>No, I read it right here, maybe about a month ago. Sorry, I don't remember
>>>who's doing it, maybe somebody else does. I hope the results will be posted
>>>here or published.
>>>
>>>>Add an 11th ply to a program that knows nothing and
>>>>it's importance will be far greater to it than to a program that has a very
>>>> large evaluation function.
>>>
>>>Are you sure? I'm not. In fact I've seen arguments that CSTAL gains more with
>>>increased time than fast searchers.
>>
>>Not increased time, increased plies. Knowledge vs. speed. If a program knows
>>nothing except the material values, then it can only reside on tactics as it
>>depends on its calculations to reach any conclusive decision. In other words it
>>must actually see a side losing or winning something to calculate an advantage,
>>meaning that most of its decisions are completely random, right? Now add just
>>one element of knowledge and provided the knowledge is correctly balanced with
>>the material values, the decision will become that much less random for the
>>positons where the knowledge is applied. This becomes increasingly true for each
>>and every scrap of correct and properly balanced knowledge placed in the eval
>>function. Now, let's get back to that extra ply (not time). Our program
>>MindBlank has calculated 10 plies, and thus requires every ply it can get to try
>>to make its decision that much less random. On the other hand we have CyberGM
>>who truly knows what's going on, and at 10 plies already has a solid decision on
>>what the next move should be and whose 11th ply, short of showing a forced
>>material win. That's why more knowledge is the only way to go, but it must be
>>done carefully or it can be just as much of a problem as the solution it is
>>supposed to provide.
>>
>> Albert Silver
>
>The difficulty is that by the time CyberGM reaches 10 ply, MindBlank will be way
>past 10 ply.
I quite agree, but that's why I rectified his statement of a smart program (such
as CS-Tal) benefitting from more time, as the issue was the Elo value of a ply
not the ELo value of a minute.
> Null-window searches on material-only would smoke right along.
>Your example is extreme enough that I would still expect CyberGM to play better
>chess; I am writing to point out that a generalization will not necessary hold.
>
>Even your conjecture, based on ply, is up for some debate. Something that
>doesn't know much might improve more from an extra ply than something that knows
>quite a bit. MindBlank at the same search depth as CyberGM will play much
>worse. As search depth is increased for both, both programs improve, but
>MindBlank may improve more quickly, simply because there is more room for
>improvement!
That's exactly what I said. With no knowledge other than a material count, the
only 'understanding' a ply can bring is a forced material win for one side or
the other. Which means that for every line that doesn't achieve this, the
results are quite simply random.
Example: After 10 plies, I still have no forced material win for any side, thus
all my lines have the same evaluation, +0.00. Right? Any choice I make among
these lines is utterly random, as the only criteria I have to choose between
them (material) doesn't do the trick. Therefore, an 11th ply is my only hope for
a better than random decision. That was the example of MindBlank.
Now I have CyberGM. Here, I have a program with a very large amount of
information, so that after 10 plies, it already have a very precise idea of what
to do, and have chosen move 'a' as opposed to moves 'b' and 'c' based on its
knowledge. An 11th ply can most certainly help improve the precision (depending
on how precisely implemented the knowledge is) of the program, but isn't
_necessary_ to find a reasonable move. Each added element of knowledge makes the
decisions at each ply that much less random, so that it is not dependant on the
benefits of a forced material win that an extra ply might bring it. Sure, the
11th ply might bring about a material win, but otherwise it's benefit will
likely be less than the +1.00 required by MindBlank to make a decision.
Albert Silver
>
>Dave
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.