Arimaa Forum - Print Page


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

        Arimaa >> Say Hello >> Work Avoidance Project ;)
        
(Message started by: hrb on Nov 15^th, 2012, 2:31pm)

Title: Work Avoidance Project ;)
Post by hrb on Nov 15^th, 2012, 2:31pm

Hi guys,

I've read all the http://arimaa.com/arimaa/papers/ and I've decided that I want to create a bot. IINM, Omar intended Arimaa to stimulate the development of general AI, so someone should take a stab at making bot AI more general. IMO, at present almost all improvement in bot playing strength is due to the intelligence of the human programmers, i.e. learning from experience and changing the software accordingly. In contrast, an intelligent bot should really be able to learn from scratch that a game without horses is identical to a game without cats, if and only if it is worth the effort.

Heading to the bot_development forum...

Title: Re: Work Avoidance Project ;)
Post by dht on Jul 31^st, 2013, 1:31pm

on 11/15/12 at 14:31:30, hrb wrote:

Good luck :D

Title: Re: Work Avoidance Project ;)
Post by hrb on Jul 31^st, 2013, 2:09pm

Thank you!

Unfortunately, I got sidetracked by stupid real work and i spent most of my free time napping.

To make things worse I just recently bought the most awesome car ever ('04 Mazda RX-8), which I decided would be even moster awesomester if it had a turbo and 400 rwhp. ;)

Title: Re: Work Avoidance Project ;)
Post by dht on Jul 31^st, 2013, 2:23pm

on 07/31/13 at 14:09:27, hrb wrote:

Do a quick overnight hack job with an feedforward backpropagating neural network evaluating all possible moves generated from the current position, and leave it in the basement training against itself for a while then go check on it?

Title: Re: Work Avoidance Project ;)
Post by dht on Jul 31^st, 2013, 2:46pm

Actually when I was experimenting with that sorta thing, I found 2 competing neural networks training against each other one for noughts and one for crosses training against each other with intermittent play against a human had the most promise

Didn't play that strong, but once it beat me on a 10x10 board with 5 in a line to win by showing some creativity and misdirecting me developing one area early on then shifting play to a different area and then connecting from the area we were both competing in to the dead area...

It wasn't too strong, but it was cheeky and had personality as well as showing signs of a bit of psychological play and it was d**n creative

I used http://leenissen.dk/fann/wp/ and coding it up took me around a day at most.... No optimizations or encoding the board, just a set of input neurons for the board, empty at 0, -1 for noughts 1 for crosses (or the otherway around it was over 6 years ago and I lost the code in a hard drive failure) tanh activation function and the network trained with an evaluation function such that pr win and loss over time were just 1/x with x being 1 for the winning position and incrementing? each earlier step so it was just 1/1 for winning position 1/2 for the next or something like that, basically something looking a bit like log back to front and upside down...

What gave it the sparkt was separating naughts. for arimaa that would be gold, and crosses (silver)

Self playing NNs that just played against themselves, were alright, but boring

Give it a go? ::)