Arimaa Forum - Print Page


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

        Arimaa >> Bot Development >> let the machine write code?
        
(Message started by: clauchau on Sep 4^th, 2004, 10:30am)

Title: let the machine write code?
Post by clauchau on Sep 4^th, 2004, 10:30am

It seems a good idea to check for goal reaching with fast if-statements and dedicated flags and possibly elementary routines, as Bomb and Speedy seem to do. Writing them by hand is possible since Dave did so, but to me it looks already tedious for a 4-steps check. Adding 4 counter steps and 4 further helping steps seems too much work for me :P if I keep in mind the goal to improve Artificial Intelligence.

So, how about making a program write it for us? Might not be a perfect test and it might well span one Mega bytes of brute and redundant code, but it could be made to evolve and get shorter? It could start from a million random board positions and try to see some simple pattern, with some gentle human assistance.

Hmm, it's always easy to throw ideas and I'm not sure how practical this one can ever be.

Title: Re: let the machine write code?
Post by 99of9 on Sep 5^th, 2004, 8:51am

That sounds even harder than writing an arimaa bot in the first place!

From what i've heard, genetic programming has not had that much success at useful applications. (different from genetic algorithms of course)

Title: Re: let the machine write code?
Post by clauchau on Sep 12^th, 2004, 6:22pm

I'll have a try at the rule induction system at

http://www.newty.de/pnc2/

There is a Chess Endgame benchmark to test.

Title: Re: let the machine write code?
Post by fotland on Nov 8^th, 2004, 5:04pm

It's not that bad :) Bomb has about 500 lines of code to produce a good lower bound on the number of unopposed steps needed for a rabbit to reach the goal (range 1 to 8, or "more").

I try to statically evaluate counters to the goal threats. Instead I use the larger values to drive search extensions with limited move generation.

It took about 2 months to write and debug. I had planned to work on trap/frame/pin evaluation next, but ran out of time last year.