Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: Will Sufficient Extra Knowledge Make Up For Loss Of Speed Eventually?

Author: Robert Hyatt

Date: 05:29:43 07/20/00

Go up one level in this thread


On July 19, 2000 at 11:53:14, Graham Laight wrote:

>On July 19, 2000 at 10:47:43, Ed Schröder wrote:
>
>>On July 19, 2000 at 10:22:20, Graham Laight wrote:
>>
>>>On July 19, 2000 at 10:10:04, Ed Schröder wrote:
>>>
>>>>On July 19, 2000 at 08:40:12, Graham Laight wrote:
>>>>
>>>>>On July 18, 2000 at 18:17:45, Robert Hyatt wrote:
>>>>>
>>>>>>We have several copies of Junior (and others) running on ICC, including more
>>>>>>than one deep junior.  I have seen GM players achieve these kinds of attacks
>>>>>>in very fast games...  Where the GM has little time to think and has to
>>>>>>'intuit' everything.  And intuit they do...
>>>>>>
>>>>>>There are many positions search won't solve.  There are many positions that
>>>>>>evaluation won't solve.  There is room for both in a chess engine, and _both_
>>>>>>are important when playing players at the top level of chess...
>>>>>
>>>>>I don't think that this is proven.
>>>>>
>>>>>Programmers have historically found that they reap greater dividends with speed
>>>>>than they do with knowledge, so they have mainly been going down the speed
>>>>>route.
>>>>
>>>>There is a lot of truth in that.
>>>>
>>>>>However, if, instead of going down the speed route, the same amount of effort
>>>>>and learning had gone into the knowledge route (e.g. learning how to build large
>>>>>quantities of knowledge in a systematic and maintainable way), it may be that
>>>>>knowledge based programs would now be just as strong as speed based programs.
>>>>
>>>>If speed (depth) wasn't such dominant we now would have had more intelligent
>>>>programs searching 2-3 plies less deep. One might wonder which approach would
>>>>be superior in hard elo. My guess is it is search.
>>>>
>>>>Ed
>>>
>>>This might be true. On the other hand, it may be that, beyond a certain
>>>threshold, adding knowledge increases elo at a faster rate - exponential growth
>>>if you like.
>>
>>Maybe.
>>
>>>Even if the growth in ELO as you added knowledge was only arithmetic (= steady)
>>>rather than exponential, this would (if it were true...), be better than search,
>>>where the growth in ELO rating is, roughly, logarithmic.
>>>
>>>I think we know a great deal about the effects of adding speed and nodes per
>>>second.
>>
>>After 20 years I still don't know. It's too complex at least for my brain.
>>
>>>Unfortunately, I don't think we know so much about adding knowledge to the eval.
>>
>>TRUE!
>>
>>In Rebel I have a flexible parameter called [Chess Knowledge = 100]. Setting
>>this value to its maximum (500) Rebel knows a lot more than the default
>>setting (100). Setting the paramater to 500 slows down the progranm with a
>>factor of 4-7 and it definitely perform worse. Setting this value to 25 so
>>far seems to be the best setting as it makes the program much faster. The
>>classic speed_vs_knowledge dilemma in a nutshell.
>
>Sometimes, you need to look at extremes to work out what's really happening.

I think there is already ample evidence of this.  IE a program like hiarcs,
which is very slow, vs a program like fritz, which is very fast.  They are
pretty much equal overall.  There are extremes (a program _too_ slow will
make tactical blunders, a program _too_ fast will make positional blunders)
but there is a fat "sweet spot" that is quite big (ie a factor of 10 wide,
so that 200K vs 20K is fairly equal assuming both are done reasonably well).




>
>For example, if you look at a graph of drinking alcohol VS cases of cirrhosis of
>the liver by country in Europe, there is apparently no clear trend - until you
>add France, which, by sitting well to the top right (drink lots of alcohol, get
>lots of the disease), clearly shows the strong correlation.
>
>In this instance, the case that strongly shows the trend is the human chess
>player. Lots of knowledge, but, by comparison with computers, negligable NPS.
>
>This leads us to 3 possibilities:
>
>* Humans have super-huge levels of knowledge - far more than we had realised

I don't believe this is as likely as the hypothesis that humans simply use what
knowledge they have _better_...




>
>* The elo benefits of extra knowledge are not logarithmic, as the benefits of
>extra speed are
>
>* There's something entirely different going on in the human brain which we
>haven't considered here


That is almost a certainty.




>
>-g
>
>>Ed
>>
>>>-g



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.