Computer Chess Club Archives


Search

Terms

Messages

Subject: Neural networks in the year 2400

Author: Vincent Diepeveen

Date: 18:56:53 07/04/03

Go up one level in this thread


On July 03, 2003 at 19:41:57, Robert Hyatt wrote:

>On July 02, 2003 at 12:35:38, Vincent Diepeveen wrote:
>
>>On July 01, 2003 at 13:32:19, Ralph Stoesser wrote:
>>
>>>Hello *,
>>>
>>>Why no top engine uses neural networks for positional evaluation in non-tactical
>>>situations? Are there interesting publications about neural networks and chess
>>>programming?
>>>
>>>Ralph
>>
>>because
>>  a) NN are too slow
>
>20 years ago chess programmers were saying this about _all_ high-level
>languages, as they wrote in assembly.
>
>>  b) they do not work very well for situations they are not trained for
>>     and in chess you always explore new positions which are not trained yet,
>>     which is an easy thing to understand once you understand that chess has
>>     10^44 positions and you could train perhaps for 10^2 positions at
>>     most very well so missing around 10^40 somewhere.
>
>That's a training issue.  It isn't unsolvable.

I keep hearing in my ears the dissappointment of the reporter after asking Jaap
v/d Herik (apart from professor in computer science also has a professorchair
Computer Law) when we finally would see those fantastic stories about computers
speaking law reality.
  "I expect that the first computer speaking law will be there at around the
year 2110 and by 2150 they already will be working fully automated"

After reasking whether he meant 2010 or 2110, Jaap confirmed that he indeed
meant a date long after his death.

That's what i keep hearing with (A)NNs too.

Every person will be easy figuring out that when non-random training will be
applied to teach a neural network to play chess, that the training will require
about 10^120 operations to train. 1 operation is pretty complex however. It
involves not only a tuning but also testing a big number of positions.

Therefore picking a date long after Jaap's computer speak law, is a safe guess
when ANNs will be capable of solving chess. Can we agree upon the year 2200?

Or do you guess that 10^120 will be reached much sooner by a brilliant
innovention in the year 2080? By then i will be turning 107 years, and knowing
one of my grandfathers made it to 101 years old, so there is a very remote
possibility that i will be able to confront your family by then with this
statement.

Therefore i advice you to do the same like JAAP and bet on somewhere in 2100.

>
>>  c) the persons that say they work for similar situations are on drugs
>
>I can show you an aircraft tracking ANN that worked just fine.  Very complex
>problem tracking multiple targets from one radar image to the next.

I know 1900 players on steroids that can beat fide masters...

Wait... ...you didn't say that your ANN on steroids beats hand tuned software.
I have to give you that. You're improving Bob!

If you next time claim the opteron to be 70 bits because additionally to being a
64 bits processor it's also setting a few flags, then perhaps we would be longer
enjoying Kerrigan's cool new wordings for you, as it will take him more time
then to convince even the utmost idiots here that you are already swallowing
alzheimer drugs.

Most important is that you 'forgot' to mention that airplanes always should show
the same shape at radar, which is exactly the case with voices too and which is
exactly what has been proven that NNs can do. They can, when trained well
recognize something that is 100% similar.

However in chess that happens to be not the case. In chess you have the problem
that every position that gets searched is different from what it is trained for.

So that's why this statement on steroids from you doesn't proof anything about
NNs in chess.

>>  d) training for chess takes more time than solving chess brute force costs
>>     In fact my approximation is 10^120 to train for chess a NN, under
>>     the condition that the NN has all the relevant knowledge. That is quite
>>     a big problem when you consider chess is x.10^43 according to
>>     latest findings.
>
>
>Then humans can't play chess either.  Because we have the same sort of NN
>training problem.  bottom line is that exhaustive training either isn't
>required, or else humans can't play chess.  One or the other _must_ be
>true.  And since humans do play the game well...

You are assuming that NN is using the same technology like the human brain is
build up from. Even that wouldn't be a problem, if they would more or less have
the same functionality. However also that is not the case as research has
pointed out already very clearly in the 80s and 90s. The simple model of a
braincell somewhere in the 50s was simply dead wrong. ANNs still suffer from
that problem. Go ask a brain surgeon for a more recent model of what a braincell
does.

Additionally we have a few billion brain cells and i *wonder* how you plan to
train an ANN as big as *that*.

Let's say the year 2400 for a 2 billion neuron multilayer with loops and central
collection and refeeding outputs to the input ANN?

Best regards,
Vincent




This page took 0.01 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.