[From: C. Love (920522.1200)]
[To: B. Powers (920521)]
Bill is describing transport lag to me and says,
To implement a true transport lag, you need to store consecutive values of
input in a buffer. At each dt, the contents of the buffer are shifted and
the new input is inserted at the beginning. The output is taken from the
last (oldest) entry in the buffer.
Ok Bill. Sorry about the mistake. It is a easy change to make in the
code. I am just wondering about\the smoothness of the output. I have passed
the weighted percepts and references both through a sigmoid (+- 0.5 range)
and the error signal through another sigmoid (+- 1.0). This gives me a
dynamic range for the reference output of about +- 0.5. Do you feel that it
would be wiser to subtract the weighted percept sum from the weighted
reference sum directly without using sigmoids directly on these summed
inputs? Granted the percept output would have to be sigmoided somewhere
along the way before it was passed along to the higher elementary control
modules (ECMs).
In your earlier (1979) paper you say in the Fig. 14 caption,
These signals are given quantitative weightings by the S matrix and
>summed in the input function FNI of the system to create the perceptual
>input.
>The error signal is amplified and smoothed by the output function FNO
with the result....
So does FNI and FNO use some sort of nonlinearity, i.e., sigmoids? I know
you treat the reference weights differently, i.e., {-1,0,+1}. Do you sigmoid
the internal reference signal though before subtracting it from the internal
percept signal? From fig. 14, it does not appear you do this? If you do it
for
the percepts then why not do it for the references?
I was talking to one of my colleagues here and he suggested that it may be
wise to keep the temporal sum idea while still keeping this pure delay you
state is necessary. He suggested that if I had a buffer of length 4, then for
the first three trials null output would be transmitted to the reference
output pin, but on the fourth shift the temporal summation would "kick-in".
The reason I want this is because it smooths the output , which is what I
think you meant when you said,
To prevent the oscillations and still maintain high loop gain, you have to
use a smoothing filter (running average) that keeps the feedback from
having a loop gain of 1 or greater at frequencies at or above the critical
frequency where the 180-degree phase shift occurs.
Am I correct?
In terms of progress, as you may surmise I have created the necessary
routines to create, connect, and evaluate the elementary control system
(ECS). Things move quite well when developing in Prograph. I highly
recommend it to anyone who has to develop something on this scale.
With everyone's advice I am trying to determine a good "learning" algorithm
to use, which will serve my purpose. So if anyone does have some
suggestions, don't hesitate - let me know. I appreciate all.
I will keep everyone posted of my progress. By the way Bill, I really can see
the usefulness of your "put up your feet and teach the baby" routine that uses
the grid system. I like.
Thanks again for everyone's help,
Chris Love.