[Hans Blom, 950511c]
(Avery Andrews 950511)
Thanks for a beautiful story!
One thing that differentiates a model-X perception from a real X
perception is that you don't settle for model-X when you can get
real-X ...
This is something that my demo does not demonstrate explicitly, but
some experiments are possible. Sometimes the model parameters get
stuck, for a relatively long time, at values that do not correspond
with those of the world (Bill Leach's examples of the psychotic and
the acid head). Yet control does not seem to deteriorate much in any
obvious way. The funny thing is that this condition does not arise
(as often) if some disturbance, e.g. random noise, is added to u, the
variable that corresponds with the action. This random noise
corresponds with experimentation/exploration. It keeps our internal
model better in line with what is out there.
Why? A better model allows better control. Experimentation provides a
better model. But worse control, because you do not apply u but u
plus or minus something. And that something deteriorates the control.
So there is a compromise: experiment, but not too much.
Slogan: trust your knowledge, but not too much...
By the way, experimentation does not help if you want to control at
the next time instant only. If your only goal is to bring x to xopt
at the next time instant, the random disturbance of u only makes
matters worse. It is when you know that you have to control over some
time horizon stretching into the future that experimentation at this
moment helps. Right at this moment, control is worse, but in the
future it will be better thanks to a better model.
Expectations: The young experiment much more than the old. The
internal models of the young are much more flexible/less rigid than
those of the old. The young can be taught almost anything, the old
almost nothing.
Another property of model-X perceptions is that they are defeasible,
and when shown to be false, they cease to be controlled for ...
This part is implemented in the demo by a test that checks whether
the prediction and the observation conform and, if not, resets the
model so that it can adjust to the new circumstances.
Sorry to spoil your story with more control theory ;).
Greetings,
Hans