Arimaa Forum - Print Page


    
      
        Arimaa Forum
        (http://arimaa.com/arimaa/forum/cgi/YaBB.cgi)
      

        Arimaa >> Bot Development >> bot_Gnobot_2005
        
(Message started by: 99of9 on Nov 2^nd, 2004, 1:13pm)

Title: bot_Gnobot_2005
Post by 99of9 on Nov 2^nd, 2004, 1:13pm

Just thought I'd update you on my botwriting news:

Some of you know that Gnobot / Firsttry keeps a record of information from its games in order to tune its evaluation parameters (by simulated annealing after a number of games). However this hasn't shown much improvement over the hand-picked constants I originally fed in.

So now I've written a program which goes through Omar's database and looks at every position that has ever occurred in a game between 1700+ players, and records information about that instead. The bad news is that there are a mighty lot of positions to look at, so I can't look at them in much depth (2 ply / 8 steps), and it's going to run for days before I can even start testing it out! Once I have this information I can try simulated annealing again, and hopefully might get some better constants. It's a bootstrappable procedure, so in theory if it works well once I can do the whole thing again and get even better constants...

Unfortunately constants aren't the be-all and end-all. There are many aspects of arimaa that simply are not checked for in my evaluation function, and either have to be found during the search, or I need to code them.

Title: Re: bot_Gnobot_2005
Post by omar on Nov 2^nd, 2004, 10:21pm

Glad to see that the games database is being put to good use :-)

Im also glad to see that you are using an automated approach to developing the evaluation function. In the long run I think this will prevail. In fact I hope this is the kind of approach that will acheives the Arimaa challenge someday.

I really think that our off-the-shelf computer already have the compute power to acheive the Arimaa challenge, if only we had the right evaluation function. Developing such an evaluation function by manually tweaking the parameters I don't think will be possible. I think an automated approach will be needed. However these approaches are very compute intensive.

Title: Re: bot_Gnobot_2005
Post by MrBrain on Nov 3^rd, 2004, 3:26pm

I predict that this approach will not work well. (Not to say that we shouldn't try.) Jonathan Schaeffer (of Chinook fame) tried the same approach years ago, and they found that they could get the computer to solve particular problem positions rather well using this approach. But when they played it against human-tuned parameters, it was a disaster for the computer-tuned paramemter version.

Speaking of Chinook, that team is making good progress towards solving the game of checkers. Check out: http://games.cs.ualberta.ca/~chinook/

Title: Re: bot_Gnobot_2005
Post by 99of9 on Nov 4^th, 2004, 5:05am

Can you elaborate on "the same approach"? Is there a paper/website somewhere?

I have read a number of articles from the UA games group, but do not remember anything closely resembling my method. If there is a technical paper out there then I can share the differences in my approach and why I think it may work.

Nevertheless, it's great to have equal and opposite predictions from the two of you. Now we can find out what happens. That's what science is about! :-)

Title: Re: bot_Gnobot_2005
Post by MrBrain on Nov 4^th, 2004, 7:33am

It's in the book "One Jump Ahead" by Jonathan Schaeffer. It may not be exactly the same approach (I don't know the details of what you're doing, nor his), but it's at least similar. Anyone here who hasn't read the book should immediately get a copy. Outstanding book on programming computers to play a strategy game.