[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [computer-go] ANN(s) playing a perfect game...



Campbell,

You can find one part of what I am working on, the part I have matured enough to release under the GPL:
http://sourceforge.org/projects/semiann

It is my first public iteration. I did several private iterations. It is optimized for the Sun and IBM JVM JIT(HotSpot) compilers. If you have any comments on it, let me know. I have had other priorities come up that have taken me away from this project for a month. I hope to return in early January. 'Tis the season of many distractions.

As I progress through parts and mature them for public consumption, I will release them under the GPL on sourceforge. No point in having people reinvent the low level crap, at least if they are working in Java (regardless of Diep's closed view of the world). It works better if more of us are working at higher levels on actual interesting problems as opposed to lower level optimizations. I am convinced the Go breakthrough(s) are all at way higher levels than the myopic discussions about tactics that seem to predominate discussions (mainly because that is what is more readily achievable given the strong influence of traditional Boolean methods).

I will likely hop on board David Weiss's Moyoman once I am preparing a client to play in tournaments, etc.. I have no interest in UIs, Go comm protocols or any of the low level boring rote programming crap. I am very thankful David has chosen to handle all of that for me.


Jim O'Flaherty


Campbell Macdonald wrote:

Jim:

I enjoyed your note. Do you have a site where your
ANN work is published? Let me know. If this is
proprietary, I understand.

Thanks
Campbell


--- "Jim O'Flaherty, Jr." <jim_oflaherty_jr@xxxxxxxxxxxxxxxxx>
wrote:


Imran,

I am working with ANNs using GAs. What I am finding
is there is a huge number of people who are out of synch with more
modern ANN techniques, and write ANNs off out of hand. It is not unique to
this group. There is enormous amounts of information in a number of
domains. And there even more "less informed self-righteous" software
developers.

No one (including the uh..."geniuses" of this group)
has the time to follow all the tangents in the many differing
related fields. And no single technique I have seen discussed here is going
to "solve the computer-go problem" such that the program based on
their particular "uber" technique will achieve a dan level rating
against similarly skilled humans, at least not consistently and
long-term. If hardware speed keeps doubling, perhaps it might happen in 10
years.

I suspect a hybrid approach, which a couple people
have referred to, will be what finally makes a substantial
breakthrough. Given the type of dialog I see in this group, I am very suspicious
it will come from here. I personally detect very little spirit of
collaboration. And I think the hybrid is going to require that. I do see
a member or two trying for collaboration. And most of the rest of
the group is so scarcity (it's mine, it's a secret, I'm brilliant
and won't share my *ultimate* partial solution with you, etc.) focused,
it is very disappointing. It does not generate the social
cohesion I think will be required for the hybrid. Still, I am getting some
value from the conversation. And I am getting a good idea of the
people I think I will find interesting in working with later.

And as you said, it borders on trivial to get an ANN
to perform near perfect on the simpler games. I know as I have done
it, been on a team that has done it or dl'ed and tried it out for
tic-tac-toe, connct-four, gomoku/pente and other small games. As to ANNs
playing checkers, chess, Go, etc. with perfect play, HA! I suspect that is
decades off without some parallel hardware breakthrough. And even then,
it smacks of "trying hard" as opposed to "trying smart". And it
will happen much sooner for checkers/draughts than for chess. And it
will be eons after it happens for chess before it will happen for Go. In fact, perfect play is likely unprovable ideal for Go. I had put
impossible - but with proper encoding on a future quantum computer, it
might actually become achievable. Who knows?


Jim O'Flaherty


Imran Ghory wrote:


On Sat, 4 Dec 2004, John Tromp wrote:




Imran Ghory wrote:



Nine men's morris and Connect 4 however both are,

and both can be played

to a very high (better than human) level by

neural networks.



You mean better than the average human, right?
To beat the best humans at Connect-4 requires

(near-)perfect play,

which appears quite beyond the capabilities of

neural nets.



nope, neural nets can reach near-perfect play. Just

knock up a quick test

using tic-tac-toe or some other similar game and

see for yourself.

Imran



_______________________________________________
computer-go mailing list
computer-go@xxxxxxxxxxxxxxxxx


http://www.computer-go.org/mailman/listinfo/computer-go/





__________________________________ Do you Yahoo!? Yahoo! Mail - 250MB free storage. Do more. Manage less. http://info.mail.yahoo.com/mail_250
_______________________________________________
computer-go mailing list
computer-go@xxxxxxxxxxxxxxxxx
http://www.computer-go.org/mailman/listinfo/computer-go/


_______________________________________________
computer-go mailing list
computer-go@xxxxxxxxxxxxxxxxx
http://www.computer-go.org/mailman/listinfo/computer-go/