Re: [computer-go] how to use GTP in place of GMP

[Nick Wedd <nick@xxxxxxxxxxxxx>]

In message <200408112122.i7BLMwB03065@xxxxxxxxxxxxxxxxx>, Don Dailey 
<drd@xxxxxxxxxxxxxxxxx> writes
> Heikki,
>
> I  think manual  play is  a good  idea.  It  might seem  to be  a step
> backwards, but the human element is hugely enhanced by this.
>
> Years ago in computer chess it used to be that all the program authors
> met for a tournament and you  sat down with your opponent and operated
> manually.  You chatted the whole  time and enjoyed the experience.  In
> later years, you  would often be sitting opposite  some Joe Shmoe that
> didn't know anything about the  program, the real author couldn't make
> it, he  was just  an operater provided  by the  tournament organizers.
> That wan't  much fun.   I could  have stayed home  and let  some "Joe"
> operate my  program.  I can't see it  being nearly as fun  to sit back
> while the computers are silently working in the background.
I would like to give GTP a rest for a while, and have time to absorb the 
ideas which people have presented in this thread.  Meanwhile, I will 
prattle about computer Go events.

When two programmers, with their programs both claiming to support GMP, 
get together, three things can happen.  Sometimes the programs 
communicate correctly, and play a whole game.  Sometimes they play part 
of a game, until the comms stops working for no obvious reason;  then 
the programmers simply copy the rest of the moves from one machine to 
the other.  And sometimes the game never gets started, and after a few 
tries they give up and do all the comms manually.  I have never regarded 
this as a problem.

If I really wanted to know which was the best Go program, I would get 
hold of the leading candidates, persuade the programmers to support some 
comms protocol, and leave them running overnight playing against each 
other, with a script to play game after game.  After two weeks I could 
publish a statistically significant result.  And no-one would be much 
interested.  I am glad that real computer Go events are not like that.

People don't attend computer Go tournaments to find which program is 
strongest.  Flying a person to another continent, and then playing just 
one game against each rival program, would not be a sensible way to do 
this.  The main purpose of a computer Go event is social.  Programmers 
who know each others' work, get to meet face to face.

Having to do manual move transfer rather adds to this, I feel.  You get 
more emotionally involved in a game if you are forced to take notice of 
every move.

The programmers are generally much stronger than their programs, and 
this gives the event an atmosphere like that of snail-racing.  You spend 
a year training your snail, and then take it along to race against the 
other snails, cheering it on while it blunders about.  Maybe you see 
your program building an implausibly large moyo, and you know that its 
opponent ought to invade;  but will it?  You can get really involved.

Some events specify a comms method, and require the contestants to 
support it.  I don't thing this is a good idea.  Having a comms method 
that works helps the contestants, so if they don't have one that's their 
own problem.  It does not generally help the organiser much.

I can think of two cases where working comms would help significantly. 
One is if the format of the event won't fit into the time available 
without having some of the programs play more than one game at the same 
time.  Normally most programs play well within the time limits, so there 
are only one or two that need to be multi-tasked.  This happened at 
Dublin in 2001.  For a program that is playing two games at once (on 
different computers of course), there are various ways you can do it. 
You can have the programmer doing the manual move transfer for both 
games;  but he isn't going to be happy about this, particularly as his 
program may be in danger of losing on time anyway.  You can use a 
volunteer to do the manual transfer for one of them.  Or you can have a 
comms method that works.  I think I prefer the last of these.

The other reason for really wanting a comms method that works is if you 
have invited programs to compete without their creators being present. 
We did this in London in 1998.  None of the entrants that arrived 
without their creator had a working comms protocol, and they were 
operated by volunteers.  I won't be doing this again - it is a dull job 
for the volunteers, who lack the emotional involvement that the creators 
would have.  I would accept programs without their creators, but on the 
understanding that if the tournament staff couldn't get a program's 
comms working within a reasonable time, then they would give up and let 
it lose by default.

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