[pdp-discuss] LSTM/RBP

[Randall C. O'Reilly]

This is not an inappropriate usage.  LSTM is difficult to setup correctly, but 
RBP should work without too much difficulty -- it does require attention to 
the appropriate architecture and parameters.  You might want to just start 
with an SRN (just add a context layer to a feedforward backprop net -- the 
wizard will do this for you) -- it is the most foolproof and will work pretty 
well, but not as well as the others.  This will give you a quick first-pass 
sense of whether there is any regularity at all.

- Randy

On Monday 05 December 2005 08:46 pm, Dmitry Kit wrote:
> Hello everyone,
>
> I have the following problem. I have resource usage trace of a program and
> want to see if I can predict the next usage level. The numbers are all
> between 0 and 1. I was thinking a recurrent neural net would be ideal for
> this situation. Given a current usage pattern give the next pattern as the
> output target. What I see however is the net quickly reaching a steady
> state where some units are always on and others are always off.
>
> The question is this even a good use of recurrent nets? I know it works
> with n inputs activated in a certain pattern (0 or 1). In my case I only
> have one input and the activation is fairly close together, although I
> don't see why the output settles on a high activation value (1). I don't
> want to spend too much time on this if I'm using it incorrectly.
>
>          Best Regards,
>                        Dmitry