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

"Eternal Vigilance" <[EMAIL PROTECTED]> wrote in message
news:[EMAIL PROTECTED]
> AngleWyrm wrote:
> > I seem to recall that Learning is opposite Memorization; a balance
between
> > flexibility and rigid precision. What if we let the system continue to
> > learn? It is my guess that such a system would 'forget' old information
as
> > new information continued to update the weights, which seems to mirror
> > nature.
>
> Neural Nets are notorious for becoming unstable if overtaught (and even
> failing alot to form properly in the first place, requiring many
repetitions just to create
> a NN solution that 'works'.   The usual back propagation learning requires
the entire
> training set to be applied consistantly.  If you try to dump in a new
subset onto a previously
> formed NN it will destabalize the patterns formed by the previous data
set.

Destabilizing the patterns formed by the previous dataset seems to me to be
exactly the process of forgetting: Replacing old information with new
information...I begin to wonder what will be forgotten and what will be
remembered.

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?

Anyone know of some downloadable tools I could experiment with?