Re: Commercial video game that learns from its own mistakes?

"AngleWyrm" <[EMAIL PROTECTED]> wrote:
> Which brings up a point: Is it possible for such networks to continue
> learning? Does the word Traditional apply to these algorithms?

OK, baylor's little rant on neural networks

Neural networks is a stupid term and everyone who knows better and
uses it should be shot. Or marketing people, 'cause i just plain don't
like marketing people (just kidding, i worked at Best Buy and we had
some nice marketing types)

Neural network is the "doesn't that sound cool?" term for any
algorithm that uses linked lists. And the nodes in the list might have
to be floating point numbers (not sure). That's pretty much it

Traditionally, ANN referred to a feed-forward back-propagated network.
You had a few nodes that took floats (0-1) linked to some other nodes
that had no value linked to the output nodes (normally one, could be
more if you're doing fixed-set classification, which is all a normal
ANN can do). And the output node is normally 0-1. Actually, almost
every node except the inputs is 0 or 1, not really a number inbetween.
It's what happens when you use a sigmoid transformation function. As
my teacher said when i showed how an ANN was just a function where the
constants are stored in a linked list, if you aren't using sigmoids it
probably isn't/shouldn't be a NN

So you plug in a number b/w 0-1. It gets multiplied by some number
stored with the link to the next node and the result is stored in that
node. It's linked to the output node so you multiply the node number
by another number and you have your output

If you had a NN with 1 node in each of three layers, and the "weight"
of each link is 2, plugging in .5 results in:
  0.5*2 = 1 (middle node value; the "hidden" layer)
  1*2   = 2 (output node value; the answer)
We have now written a neural network to solve x*4. Whee. Of course, in
a "Real" neural network, we'd use sigmoid functions that basically
just round numbers - 0.4 gets pushed down towards 0, 0.6 gets pushed
up towards 1, 0.5 normally stays 0.5. It's those pushes that makes the
network "fuzzy"

So a traditional feed forward NN is just a wierd way of writing a math
function (2*x, x^2+4y+1/2z, etc.) using a linked list. Given that it
represents a single, specific function, learning here just means
learning what numbers to multiply (and it's always multiply) the
inputs by to get the right answer. There's no adaptation (you don't
continue to learn) because 2+2 doesn't change (technically, neural
networks can't solve problems as simple as 2+2; it has to be a
logic-gated function that can be solved with multiplication; this
assumes you're using sigmoid functions or whatever and not simple pass
thrus)

So is it possible for an ANN to continue to learn? The question
doesn't really make sense - there's nothing to continue learning. It's
just a math function


But, if you remember that people are calling anything that uses a
linked list a neural network (and yes, it's just a name, it has no
relationship to the neural networks in the brain), then you're free to
write whatever you want

In the spirit of discovering math functions, there's the SVM (support
vector machine), which is an ANN that uses "kernel functions". You
have tons of grouping algorithms (which are nothing at all like
feed-forward nets) with names like SOM (self-organizing map) and
Boltzman machines. You have feature extractors (basically, trilinear
filtering applied to data) such as LeNet. And there's a bunch of other
stuff, all of which share the marketing name "neural networks"

So could you make something that adapted and then call it a neural
network? Sure. The question is, why would you want to

-b, the eternal optimist