Arimaa Forum - Print Page


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

        Arimaa >> Say Hello >> Hello World!
        
(Message started by: Migi on Aug 19^th, 2011, 1:01pm)

Title: Hello World!
Post by Migi on Aug 19^th, 2011, 1:01pm

Hi!

I'm Migi. About a year ago I've come across Arimaa, played some games and started making a bot, but I never really got far and quickly lost interest. It didn't take long before I realized that plain old alpha-beta search isn't going to cut it without a really good evaluation function, and I was just not good enough as a player to write such an evaluation function.

But recently I was reading about neural nets and I got an idea of a new approach that might just work for Arimaa. I tried to search on Google if something like it has been tried before, but Googling ideas is pretty hard. So I didn't find anything, and I decided to just give it a shot. I guess we'll see in about a week or two how it goes. But even if my bot turns out to be a weak opponent, it think it'll be interesting to play against. It will think a lot more like humans and much less like the other bots. Bad tactically and (hopefully) strong strategically, which is what you need for Arimaa.

But I'm trying to stay realistic here. People have tried to make neural nets play chess and that hasn't really worked out. And basically I'm trying to beat humans at their own game here. Trying to outsmart 100 billion neurons with about 1000 neurons, you have the right to call me insane ;D

In the meantime, I'm also playing Arimaa again. And I'm enjoying it a lot! The YouTube videos with commentary were great fun, and they improved my game sense a lot too. So a big thanks to everyone who made or helped to make those videos! :)

Title: Re: Hello World!
Post by Eccentric on Aug 19^th, 2011, 1:11pm

Insane? Peh, you have to start somewhere. As humans have to practice Arimaa to get good at it, programmers have to practice programming for bots to get good at Arimaa! In any case, I hope your stay here will be a lot longer than last time's.

Title: Re: Hello World!
Post by Fritzlein on Aug 19^th, 2011, 9:56pm

Did you read Haizhi Zhong's thesis? He tried to make a learning bot but it didn't play as well as his hand-coded one.

http://arimaa.com/arimaa/papers/HaizhiThesis/haizhiThesis.doc

If you make any kind of learned evaluation function that outstrips a hand-tuned evaluation, everyone will think you are cool, and Omar will be deliriously happy, since that was the whole point of the Arimaa Challenge.

Title: Re: Hello World!
Post by Migi on Sep 1^st, 2011, 5:23pm

on 08/19/11 at 21:56:35, Fritzlein wrote:

Thanks for that link. I read through most of his paper and there are quite a few good ideas in it that I didn't consider before.

However, the aspect I'm most interested in, the automated learning, has only 2 pages, in which he just says very briefly that it didn't work out. He says it may be just a bug in their code or that they didn't train it long enough, but nothing more. I would like to get in contact with the author, to try to find out more specifically what he tried (did they use a library or self-written code? what was the learning rate? etc) and what went wrong (did it learn a little/not at all/got worse by learning?). So, does anyone have any contact information about this person, Haizhi Zhong? I searched on google but didn't find anything. I would appreciate it a lot! Thanks.

Title: Re: Hello World!
Post by Fritzlein on Sep 1^st, 2011, 6:37pm

I don't have any contact info for haizhi, but it occurs to me that his thesis adviser might be willing to help you out.