OSDev.org

Hi all...

I'm taking a break from OS devving until I get my inspiration back (and end my constant fight with the compiler).

In the meantime, I'm working on artificial intelligence, mainly, neural networks. I've looked around the net and found some good documents on neural nets, but I have a few questions that I hope you guys can answer.

Firstly, in what way do the neurons communicate? The way I look at it at the moment, each neuron is trained to do something different. One example I looked at had one neuron recognizing 'T' and one recognizing 'H'. They basically output '1' if the pattern was a match.

However, this feels wrong, mainly because I thought the neurons were meant to work together to achieve a goal...

Secondly, how does the AI 'learn'? Do I just setup a linked list of 'rules' for each neuron that I teach them and then make each neuron check its input against the rules? Or am I way off track?

Finally, can anyone recommend any (free) resources that I could use to assist me?

Thanks in advance.

Hi pcmattman,

Neural networks are pretty easy in theory, but harder to implement for some of the various learning schemes.

The basic premise is a 'cell' has multiple inputs, a summing function, some weights and an output. So what happens in operation is it obviously sums the inputs, which have a weighting. If this value exceeds a threshold, it fires. That's the basic perceptron model.

On the topic of learning, there are many methods available. Supervised, unsupervised blah blah. Something you should look into is feed-forward and feed-back error propagation, a method for learning through reinforcement. There's also middleware which can help you do the dirty work.

Ideally, any starting network would have each levels nodes connected to each of the next layer's nodes with stochastically-assigned weights. But yeah, a cursory Google search will reveal all this information in greater detail.

Solidus.

Solidus117,

Thanks, I couldn't figure out the whole weightings idea until you explained it.