Neural networks
Bill Keller
Has anyone made any kind of effort to use a pure neural network for
move analysis? With the large number of games that are out there, it
seems like there ought to be enough material to train a network with.
And with processor and memory prices going down while getting more
powerful, one could be build a very large network.
My thought was 64+ input nodes (one for each square, then some control
nodes for whose turn it is, turn number, etc), a whole bunch of middle
nodes, and two output nodes (the square from and square to).
It would only search one ply deep, but it would be the RIGHT ply. :)
I plan on doing this, mostly to increase my understanding of AI and
neural networking, I'm just curious if anyone else has given this a try.
Thanks!!
--
Bill Keller
bill at arda dot net
(don't use the ReplyTo field)
Sent via Deja.com http://www.deja.com/
Share what you know. Learn what you don't.
Dr A. N. Walker
> Has anyone made any kind of effort to use a pure neural network for
> move analysis?
Yes, quite a few people. My impression is that the
results are mostly less than impressive. Chess is perhaps
too chaotic for this approach to work.
> It would only search one ply deep, but it would be the RIGHT ply. :)
My just-finished research student, Stephen Milner, trained
a neural net to analyse *positional* moves, where one ply is all
you need. We have some evidence that tacking such a net onto a
traditional search engine which looks *only* at tactics makes a
weak program [mine!] play somewhat better.
As a side-issue, the network has some interesting properties.
One of these is intransitivity of positional values -- ie the net
may prefer move A to move B, move B to move C, and move C to move
A. This cannot occur with "traditional" evaluations, which map
all moves onto ordinary numbers, but incomparable values [not quite
the same thing, but related] are well-known in game theory in
general. I have my suspicions that this is a phenomenon in human
play as well. If any of the AI people take this up as a research
topic, (a) I'd like to know about it, and (b) remember that you
saw it here first!
--
Andy Walker, School of MathSci., Univ. of Nott'm, UK.
a...@maths.nott.ac.uk
GCP
<bill_a...@my-deja.com> wrote:
>Hello all -
>
>Has anyone made any kind of effort to use a pure neural network for
>move analysis? With the large number of games that are out there, it
>seems like there ought to be enough material to train a network with.
>And with processor and memory prices going down while getting more
>powerful, one could be build a very large network.
>
>My thought was 64+ input nodes (one for each square, then some control
>nodes for whose turn it is, turn number, etc), a whole bunch of middle
>nodes, and two output nodes (the square from and square to).
>
>It would only search one ply deep, but it would be the RIGHT ply. :)
>
>I plan on doing this, mostly to increase my understanding of AI and
>neural networking, I'm just curious if anyone else has given this a try.
Work on this has been done...try searching for the KnightCap chess
program. It comes with sources.
--
GCP
Olivier Scalbert
If I have time, I want to do a positional evaluator for calm position for go
game, based on neural network. This evaluator will be used in a classic
alpha beta framework. The network's tuning will be done with genetic
algorithm staff (competition between individus (program) in each
generation). I will start on 9x9 size goban.
The biggest problem is, I HAVE NO TIME !.....
Olivier
Jorge A. Marta V.
>
> Hello all -
>
> Has anyone made any kind of effort to use a pure neural network for
> move analysis? With the large number of games that are out there, it
> seems like there ought to be enough material to train a network with.
> And with processor and memory prices going down while getting more
> powerful, one could be build a very large network.
>
> My thought was 64+ input nodes (one for each square, then some control
> nodes for whose turn it is, turn number, etc), a whole bunch of middle
> nodes, and two output nodes (the square from and square to).
>
> It would only search one ply deep, but it would be the RIGHT ply. :)
>
> I plan on doing this, mostly to increase my understanding of AI and
> neural networking, I'm just curious if anyone else has given this a try.
>
>
> --
> Bill Keller
> bill at arda dot net
> (don't use the ReplyTo field)
>
> Sent via Deja.com http://www.deja.com/
> Share what you know. Learn what you don't.
My suggestion will be to use 64 input nodes times 6 (six different
pieces options) times 2 (black or white), total 768 input nodes.
The output nodes 128, 64 (from) and 64 (two).
Good luck.
Anders Thulin
Bill Keller <bill_a...@my-deja.com> wrote:
>Has anyone made any kind of effort to use a pure neural network for
>move analysis? With the large number of games that are out there, it
>seems like there ought to be enough material to train a network with.
Assuming, of course, that the games out there are useful for such
purposes. It may be that they're not.
There is however, *complete* information on endgames. It might be
easier to start on those. It would be pretty useful as well if a
5-piece endgame database could be replaced with a (smallish) network.
Just getting a K+Q vs K network going would probably be an
interesting task.
>It would only search one ply deep, but it would be the RIGHT ply. :)
It wouldn't search anything at all. It would produce something that
would conform to the learning. If that is The Right Ply is largely a
question of how complete learning has been. (Endgames are good that
way: the entire learning space is perfectly known.)
In an endgame situation it *might* be interesting to produce more
than one output: after all more than one move may lead to the goal,
and there seems to be no reason to force the ANN to produce one move
rather than another equivalent one. It *might* complicate learning to
force the network to implement a discrimination function.
I suspect, also, that the whole question of repeated positions
should be handled outside the network.
--
Anders Thulin Anders....@telia.se 013-23 55 32
Telia ProSoft AB, Teknikringen 6, S-583 30 Linkoping, Sweden
Gary
http://home.seol.net.au/luke/Octavius/
Cnidarian
On Wed, 01 Sep 1999 17:08:52 GMT, Bill Keller
<bill_a...@my-deja.com> wrote:
>Hello all -
>
>Has anyone made any kind of effort to use a pure neural network for
>move analysis? With the large number of games that are out there, it
>seems like there ought to be enough material to train a network with.
>And with processor and memory prices going down while getting more
>powerful, one could be build a very large network.
>
>My thought was 64+ input nodes (one for each square, then some control
>nodes for whose turn it is, turn number, etc), a whole bunch of middle
>nodes, and two output nodes (the square from and square to).
>
>It would only search one ply deep, but it would be the RIGHT ply. :)
>