Re: Productive SRNs

["Randall C. O'Reilly" <oreilly@grey.colorado.edu>]

Javier =?iso-8859-1?Q?Mar=EDn?= <jms@um.es> writes:
> I'm trying to set up a trained SRN network (Elman 1990) to be productive. 
> That's to say, I want the network form a arbitrary starting input produce 
> an output, normalize this output and with a probabilistic algorithm set up 
> next input to the network from the normalized output, and continue this for 
> a number of cycles.

This can be done with a script -- you can read the output activations
into a float_RArray, which has a function for normalizing itself, then
use that as probabilities to pick a random input.  to set the next
input you should set the epoch to SEQUENTIAL and then use
epoch->trial.val+1 as an index into env->events to identify the next
event presented to the net, and then modify the input pattern of this
event.  

This would take only a relatively few lines of well chosen code.  the
problem is finding that code.  The 'type' command in css is invaluable
in identifying the functions and fields available to you, along with a
brief description.  try typing:

type float_RArray

at the prompt and you'll get a sense of what you can do.  sorry, but
that's as much help as I can give right now.

				- Randy