[From Bill Powers (950627.2050 MDT)]
Avery Andrews (950728) --
I've become skeptical of the utility of feeding x_ref into output
functions.
:What I'd suggest instead
for total error-correction against a constant disturbance is:
>
> x_ref
v
------> C ----------
> > x_err
> x_perc O O(x_err) =g*x_err + K
> >
> v
- - - - P - - - - - - - - - E - - -
Where P is the perceptual function, E the effector, and O the smart
output circuit, whose smartness is supposed to consist of a capacity to
slowly adjust K until x_err vanishes.
The simplest form of your "smart O" is simply an integrator: as long as
there is any error signal, the output E keeps changing. If you want fast
response, too, you can add a proportional component to the output
function.
Of course your suggestion applies only when the disturbance of P is
accurately predictable or constant.
The reason for feeding x_ref into the output function as well as the
comparator was to provide a simple way to continue producing output open
loop when the input is lost. If the error signal can be set to zero when
the perceptual input disappears (as it can be using balanced one-way
comparators), the output will become just what it would be with no
disturbances (given suitable weightings here and there).
···
-----------------------------------------------------------------------
Best,
Bill P.
[Avery Andrews 950628:1500]
(Bill Powers (950627.2050 MDT)
>The reason for feeding x_ref into the output function as well as the
>comparator was to provide a simple way to continue producing output open
>loop when the input is lost. If the error signal can be set to zero when
>the perceptual input disappears (as it can be using balanced one-way
>comparators), the output will become just what it would be with no
<disturbances (given suitable weightings here and there).
Right (somehow lost view of this overnight). So one question is whether this
kind of smart output will be more useful than smart input, in complicated
hierarchical systems. If a single `loosable' perception is involved in
the perceptual functions of several control systems, it might be
better to manufacture one ersatz perception rather than smart up
several different output functions. One way of doing this would
be to monitor some aspect of the activity of the effectors that are
supposed to be affecting the controlled perception, so as to find
a function whose output will track x_perc when the real x_perc
is deemed availabel and valid, and substitute for it when it isn't:
> x_ref
v
------> C ----------
> > x_err
> x_perc O | |
S<--------------- |
^ \ |
> \ v
- - - - P - - - - - - - - - E - - -
S is the smart perceptual function; its output normally equals
the output of P, except under the exceptional circumstances where
it outputs something based on monitoring of E, where the processing
of this `proprioceptive' input is an attempt to internally model the
the relationship that the environment induces between the output of E
into the world and the output of P into system.
One reason I think there might be such things is the bizarre feeling
I get when I walk on a stopped escalator: I conjecture that the feeling is
caused by failure of the usual correspondence between visual flow
rate and locomotory activity: the feeling is basically S getting retuned.
This example also illustrates the point that a smart input function can
collect into from an arbitrarily large number of sources (& one of mine is
smart enough to detect when I'm on an escalator, but not when it's stopped!
The smart perceptual function also isn't limited to looking at driving
signals for effectors, but can pick up info about what they're actually
doing. E.g. get propeller rpm's rather than throttle-position. This ought
to confer greater robustness.
If course this talk about `real' and `ersatz' perceptions, is pretty loose,
and could use some hammering out, but I hope the idea isn't too obscure.
And that I haven't made more idiotic blunders....
Avery.Andrews@anu.edu.au