[Avery.Andrews 9303130630]
(Bill Powers (930312.0830))
>Do you have a copy of Arm Version 2.0? It allows you to set up
>two target positions, and then have the target jump back and ...
Yes, & I'll try it out, but Arm2 is quite complicated, n& I was talking
about the behavior of a very simple system. I'm not yet sure exactly
what properties of the Arm2 model allow it to produce straight lines in
this situation.
>In what sort of situation? And what is doing the pre-determining?
>In Arm v.2, the paths for maximum-speed jumps are indeed
>predetermined, but they're not precalculated -- they simply
>emerge from the time-constants of the control systems and the
>dynamical properties of the arm. The trajectories for fast
In pointing gestures. When arms are bopped towards the target, the
return back to an appropriate point on the trajectory, & then proceed
on it. No evidence for `planning' as opposed to `generation'. Looks
to me as if slowers like those between the visual kinesthetic levels
in Little Man shoudl be able to do this kind of generation.
>>(`deciding what to do is easy when you
>>know what's in front of you' is the way David Chapman put it).
>
>Oh, cute. How easy is "knowing what's in front of you?"
Hard. That that's the hard part was supposed to be one of the main
points of his book.
(Rick Marken (930312.0730))
Just a few points - no time to say more.
>What they found was simply that in some unusual cases (like
>focusing a camera or tuning a radio) a control system cannot know "what
>to do" until it does something and determines the perceptual effect of
>doing it. The fact that people can control in these situations is simply
>evidence for a hierarchy of control systems -- the very notion that Fowler
This general point is important and well taken, but it's not a good
solution to the problem of distance-from-shoulder control when the
elbow is fully extended, since grownups in any case clearly know what
to do in this situation without experimentation.
>You can
>remap the coordinates of this point all you want -- Catersian, polar,
>city block, whatever -- but it probabaly won't make a big difference in
>the performance of your model.
I have already seen this to be false - adding the polar distance to the
cartesian Jacobian model improves performance considerably. I've
also seen the resulting model get stuck in situations where polar
coordinates would get it unstuck. There might be clever things to
do with the cartesian jacobians that would help in these situation
(whence the `naive' in my title), but on the whole, though I can't
think of an good reason to continue developing this approach, when
well chosen perceptual dimensions eliminate the problems without
further cleverness.
> PCT suggests that the
>jocobean matrix for computing output from error is the wrong approach;
>complex output computations like that are unlikely to be performed by
>the nervous system anyway, and they kind of bind you to a particular solution
>space (so that you get stuck when circumstances conspire to push you
>out of that space).
Actually, they're pretty simple, just a few sines and cosines which
neural circuits shouldn't have much trouble with. Maybe it would even
be useful to have them around for certain purposes, tho at the moment I
don't know what.
Avery.Andrews@anu.edu.au