Try training a neural network that uses recurrence, i.e. outputs at some
level are fed back in as inputs. This allows the net to train on time
dependent data. Examples of this are Jordan, Elman, or Real-Time Recurrent
Learning (RTRL) networks. You could also try a Time Delay Neural Network
(TDNN), which uses not only the current input data values (time t) but also
the previous several input sets (t-1, t-2. etc.). This allows the net to
view a window in time, outside of the immediate inputs. Standard backprop
networks are notoriously poor at time dependent data sets, unless you use
recursion or add time-delayed inputs.
"Ursus Horibilis" <firstname.lastname@example.org> wrote in message
> "Greg Heath" <email@example.com> wrote in message
> > Date: Wed, 28 NOV 2001 10:01:27 +0100
> > From: Giorgio Corani <firstname.lastname@example.org>
> > On Wed, 28 Nov 2001, Giorgio Corani wrote:
> > > Since identical rainfall (i.e. identical inputs) results in
> a very
> > > different runoff (i.e. different outputs) depending on the
> dryness of
> > > the soil, the dryness of the soil has in fact the role of a
> state variable.
> > > I think that such state variable could not be very useful
> is used as
> > > input to the system.
> > I don't agree. If you have no other way of estimating
> dryness, you should
> > introduce it as an input and perform a parameter study over
> the space of
> > all interesting combinations.
> Here's another thought. Use "soil dryness" as both an input
> and an output of the network. Modify your initial training set
> so that the (n-1)-th output "soil dryness" corresponds to the
> n-th input "soil dryness". Use your expert knowledge to
> establish the initial values. Train the net and then create
> some new soil dryness values using the trained net. Build a
> new training set using the new "soil dryness" values and
> iterate again.
> Will you ever be able to close the loop on the real network?
> Maybe or maybe not. The network, with "soil dryness" output
> wrapped to "soil dryness" input may go unstable after a few
> unsupervised iterations.
This message was posted through the fuzzy mailing list.
(1) To subscribe to this mailing list, send a message body of
"SUB FUZZY-MAIL myFirstName mySurname" to email@example.com
(2) To unsubscribe from this mailing list, send a message body of
"UNSUB FUZZY-MAIL" or "UNSUB FUZZY-MAIL firstname.lastname@example.org"
(3) To reach the human who maintains the list, send mail to
(4) WWW access and other information on Fuzzy Sets and Logic see
(5) WWW archive: http://www.dbai.tuwien.ac.at/marchives/fuzzy-mail/index.html
This archive was generated by hypermail 2b30 : Fri Jan 11 2002 - 17:35:29 MET