Google AI Challenge 2011

Discussion of anything and everything relating to chess playing software and machines.

Moderators: Harvey Williamson, bob, hgm

Forum rules
This textbox is used to restore diagrams posted with the [d] tag before the upgrade.

User avatar
Don
Posts: 5106
Joined: Tue Apr 29, 2008 2:27 pm

Re: Google AI Challenge 2011

Post by Don » Wed Oct 19, 2011 6:26 pm

Gian-Carlo Pascutto wrote:http://aichallenge.org/
Looks like a lot of fun!

User avatar
Kirill Kryukov
Posts: 492
Joined: Sun Mar 19, 2006 3:12 am

Re: Google AI Challenge 2011

Post by Kirill Kryukov » Thu Oct 20, 2011 2:26 am

Already almost 700 participants, even before the official start. I bet this year the total number will exceed the last year's 4617.

Daniel Shawul
Posts: 4055
Joined: Tue Mar 14, 2006 10:34 am
Location: Ethiopia
Contact:

Re: Google AI Challenge 2011

Post by Daniel Shawul » Tue Oct 25, 2011 8:34 pm

I experimented a bit with a bot derived from java starter package. It seems that the AI is dominated by path finding (Djkistra & A*) , despite my expectation that I would be able to use some multi-player algorithms. I don't think a regular turn-based depth first search is applicable here at all since the map is usually too big with many ants. Also I don't think speed of program matters that much (I maybe wrong). What language and algorithms do you use ?

There is a similar effort at stanford (for general games) http://games.stanford.edu/ that also provide starter packages. The games are provided at run time in declarative programming such as PROLOG. It requires a lot of effort to get up a random mover so they also provide starter packages in java. You can start coding alpha-beta right away. I am a bit hesitant to use the starter packages for one I want to learn declarative languages. The packages does too much for you it is like having Houndini as starter package for chess ;)

Gian-Carlo Pascutto
Posts: 1204
Joined: Sat Dec 13, 2008 6:00 pm
Contact:

Re: Google AI Challenge 2011

Post by Gian-Carlo Pascutto » Thu Oct 27, 2011 5:39 pm

Daniel Shawul wrote:I experimented a bit with a bot derived from java starter package. It seems that the AI is dominated by path finding (Djkistra & A*) , despite my expectation that I would be able to use some multi-player algorithms. I don't think a regular turn-based depth first search is applicable here at all since the map is usually too big with many ants. Also I don't think speed of program matters that much (I maybe wrong). What language and algorithms do you use ?
I'm writing mine in CoffeeScript, if only to learn that language and JavaScript. I was hoping to get something close to Python but with reasonable speed, but that's not really the case (it's quite a lot slower than C/C++ and not as good as Python :-)).

Game tree search on individual ants is not going to be possible. Maybe with grouping? I don't know yet.

sedicla
Posts: 149
Joined: Fri Jan 07, 2011 11:51 pm
Location: USA

Re: Google AI Challenge 2011

Post by sedicla » Fri Nov 11, 2011 4:33 pm

I'm using C. Using the path finding a*, and the strategy is basically defend my hill, attack enemy hills and look for food. The rest of ants I make them move straight then around obstacles.

Also when finding an enemy I check if i will die if I move, in this case I move back.

I'm using C. It is interesting to participate in this, as a break on the chess engine programming. Challenging but fun!

http://aichallenge.org/profile.php?user=4929

Edsel Apostol
Posts: 770
Joined: Mon Jul 17, 2006 3:53 am
Full name: Edsel Apostol
Contact:

Re: Google AI Challenge 2011

Post by Edsel Apostol » Sun Nov 13, 2011 1:33 pm

Daniel Shawul wrote:I experimented a bit with a bot derived from java starter package. It seems that the AI is dominated by path finding (Djkistra & A*) , despite my expectation that I would be able to use some multi-player algorithms. I don't think a regular turn-based depth first search is applicable here at all since the map is usually too big with many ants. Also I don't think speed of program matters that much (I maybe wrong). What language and algorithms do you use ?
I'm using C and Collaborative Diffusion. I've tried A* but it is expensive computation-wise for every ant and storing the path for the succeeding turns introduces complexity.

Ron Murawski
Posts: 397
Joined: Sun Oct 29, 2006 3:38 am
Location: Schenectady, NY
Contact:

Re: Google AI Challenge 2011

Post by Ron Murawski » Mon Nov 14, 2011 5:07 am

Gian-Carlo Pascutto wrote: I was hoping to get something close to Python but with reasonable speed
Perhaps Genie is what you're looking for. It's a language that resembles Python and transpiles into C code. It's about the speed of C++.
http://live.gnome.org/Genie/
You will need to download the Vala compiler from here
http://live.gnome.org/Vala/
or from here
http://code.google.com/p/vala-win32/downloads/list

diep
Posts: 1781
Joined: Thu Mar 09, 2006 10:54 pm
Location: The Netherlands
Contact:

Re: Google AI Challenge 2011

Post by diep » Mon Nov 14, 2011 7:49 pm

What is the first prize?

diep
Posts: 1781
Joined: Thu Mar 09, 2006 10:54 pm
Location: The Netherlands
Contact:

Re: Google AI Challenge 2011

Post by diep » Mon Nov 14, 2011 7:57 pm

Daniel Shawul wrote:I experimented a bit with a bot derived from java starter package. It seems that the AI is dominated by path finding (Djkistra & A*) , despite my expectation that I would be able to use some multi-player algorithms. I don't think a regular turn-based depth first search is applicable here at all since the map is usually too big with many ants. Also I don't think speed of program matters that much (I maybe wrong). What language and algorithms do you use ?

There is a similar effort at stanford (for general games) http://games.stanford.edu/ that also provide starter packages. The games are provided at run time in declarative programming such as PROLOG. It requires a lot of effort to get up a random mover so they also provide starter packages in java. You can start coding alpha-beta right away. I am a bit hesitant to use the starter packages for one I want to learn declarative languages. The packages does too much for you it is like having Houndini as starter package for chess ;)
the problem is comparable from evaluation viewpoint with international checkers. We programmed a base there. So for example advancing with just 1 stone is not so clever, you want to do it the 'napoleontic' manner and advance with entire base. You can use distance based heuristics there and when your base is dominating a lot of terrain that's a winning advantage.

You're no longer busy then with individual ants, just evaluating how far the base is. Only extensions needed for short term victories, they can however never dominate the search.

Advancing the base you can see as medieval warfare quite litterary.

So from strategic viewpoint this game is not so interesting as tactics from hundreds of years ago are interesting where in modern warfare things advanced a tad...

Post Reply