Sorry don''t have time to do any rhymes... I have looked at some tutors about neural nets on the net and I have managed to construct my own little net... Now I wonder, how could I aplly this to my game programming projects? Any tutorials or information about this?

It depends on what you want to do with it. Unfortunately, you didn't provide enough information in this area.

So, you've created a net. What kind? Is is pre-trained or real-time trained? What were you planning on doing with it? What type of game were you thinking of putting it.

Derek Smart
Designer/Lead Developer
The Battlecruiser Series
www.3000ad.com

Edited by - dsmart on 2/10/00 2:59:32 PM

a short shout-out... Hi mr.smart, I really enjoy you game BC3000ad, latest version.

I don''t think using NN''s for direct character control is a good idea. It''s certainly hard to implement, although not impossible (though only really used in academic research).
I think the best idea for using NN in games is to allow them to subtley adjust the tactics of the AI.
For instance, in a strategy game, once you have a large amount of the AI up and running you could play two computer opponents against each other with unweighted bias as to which kind of troops they decide to build and in what number. You''ll have to build up some hard data saying the number and type of troops built and how effective they are against another number and type of troop.
You can then train the network to bias the troop building process.
If you had a set number of tactics you could also use the same idea to decide under what circumstances each tactic should be used. i.e. they are fighting defensively so we should go for stealthy guerilla attacks.

The problem is you need to have hard data (input) and a set of actions to choose from (output) before you can begin to train a network.

Hope this gives you some ideas.

Mike

Let''s take Othello for example. How should I design my network from the ground? I thought using a Hopfield network, if I will be able to use it. I want the network *learn* how to play by the time(if it is possible with Hopfield. If not I could take another). He don''t have to play good or someting... My actual question is how to design a neural network... ?

Well with Othello, one way to go, would create a net with 64 input nodes and 64 output nodes, and two layers of fully connected intermediate nodes with 16 nodes a piece. Each of the input nodes would correspond with a spot on the board, and would receive a value -1, 0, or 1 depending on whether an enemy piece was there, no piece was there or the AI''s piece was there. Each of the output nodes would correspond to a possible move. If the move isn''t legal, tell that output node "bad node!" if it returns a value above a threshold value. And then try the output move that returns the highest value. Eventually you''ll train the net on the rules for Othello, and possibly some basic strategy.

SiCrane, that method won''t necessarily ''train'' the AI to learn any strategy. I would suggest another node to which the ''bad node'' data is stored. Then use that data in a model that, sort of stores all legal/illegal moves and the AI can use that to determine its moves. . Yes, I know. Overboard, beyond the scope of this discussion etc, etc, etc.

but then again, what do I know, right?

ps: Thanks for the shout out, Anonymous Poster. If you think its fun now, wait''ll Battlecruiser Millennium comes out. It virtually eclipses BC3K v2.0x but a wide margin. btw, *shameless plug* now shots are up on the site.

Derek Smart
Designer/Lead Developer
The Battlecruiser Series
www.3000ad.com

quote: Original post by dsmart
SiCrane, that method won''t necessarily ''train'' the AI to learn any strategy. I would suggest another node to which the ''bad node'' data is stored. Then use that data in a model that, sort of stores all legal/illegal moves and the AI can use that to determine its moves. . Yes, I know. Overboard, beyond the scope of this discussion etc, etc, etc.

Are we on the same wavelength on what the neural net is doing here? The distribution of weights within the nodes will store legal/illegal moves. That''s sort of how neural nets work. There''s no need to use external data stores of any sort here. You did understand that when I said tell the output node "bad node" I meant propagate and readjust the node weights along the two internal nodes layers? Also by training the net simply along the lines of playing legal games by using the output of it''s previous moves, you''ll train the net to prolonging the length of the game. i.e. not making any *stupid* moves. This is what I meant by basic strategy.