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

AngleWyrm wrote:

> "tihocan" <[EMAIL PROTECTED]> wrote in message
> news:[EMAIL PROTECTED]
> > > What if we set up a NN with a pool of info to memorize. Then, as new
> > > information comes along, drop old information off the tail of the
> dataset,
> > > and make a few passes at the new 'window on the present'. Would it be a
> > > fuzzier understanding, for the sake of adaptation?
> >
> > Most likely not, if you want to have a correct output on your new
> > dataset, the old weights will be modified and you will have a useless
> > output on the old dataset. Unless both have the same characteristics
> > of course.
>
> This is what I'm wondering: Can definition of "correct" be a fuzzy value? If
> we use real numbers rather than integers, it might be possible to trade
> closeness to a target value for flexibility.

NN do use float inside (or simulated float with large decimal count fractions).
It does cause a fuzzy effect in interpreting the data.

The problem with trying to adjust an existing NN with only ba few new training
sets
is that the backpropagation system works by applying the entire traing set
evenly
to shape the weights and produce an output that is  correct some large
percentage
(95+%) against all the training data.

You could do the dataset cycling (oldest events being retired, replaced by new)
but you would
have to rebuild the NN entirely (very CPU intensive)



Backproagation usually is done running the training set (often quite large)  and
adjusting the
weights to correct the NN  starting with large adjustment and then decreasing
the size of the
adjustments each subsequent cycle  (finer tuning near end).
So if you tried to lay on a few new  training events,  if you set the training
factor as large you will
disrupt the fine tuning done by the original building process. And if you try to
make corrections
using a low correction factor, the changes will not be significant enuf to have
the NN react to
the new data's inputs properly.


There are different kinds of NN though,   many allow incremental training, but
are more like
case-based reasoning and dont have as much ability to  generalize continuous
input values.

For games those types may actually be more appropriate as NN start becomming
cumbersome
to use as a solution when the problem domains complexity increases.