[Martin Taylor 2013.07.30.20.43]
[From Matti Kolu (2013.07.31.0100 CET)]
Martin Taylor 2013.07.30.11.42--
Randomly connecting neurons whose inputs ultimately derive
from sensors will result in a variety of computations of arbitrary
complexity. If actions consistently influence one of these computational
results, it could become a controlled perception, especially if controlling
it has some effect on an intrinsic variable.
I recently watched a recording of a lecture given by John Koza titled
"Automated Design Using Darwinian Evolution and Genetic Programming":
http://youtu.be/xIoytwJWJP8
His approach is probably familiar to some readers. Koza and his
colleagues have been using "Genetic Programming" to evolve, among
other things, analog circuits:
Yes, it's a long established technique, and one I used 20 years ago and still do when trying to optimize control simulations. The problem with trying to apply it to reorganization is that there has to be an externally applied criterion to determine which members of the candidate population survive to contribute to the next generation and which die off without progeny. In real evolution, the environment supplies that criterion and population members who have more descendants are more likely to have their genes survive, even if their phenotypes don't.
"WHATEVER THE FIELD of endeavor, genetic programming begins with a
primordial ooze of randomly generated trial �organisms� and a
high-level description of what function the organisms are meant to
accomplish the criteria for scoring their fitness.
Exactly my point. When we are talking about reorganization, there is no such high-level description, so this wouldn't work.
But there is a criterion, which is the maintenance of intrinsic variables within ranges compatible with life that is continued long enough to allow the propagation of genes. Generally speaking, the more perceptions can be controlled, the more likely it is that some of them, or some functions of some of them, affect the values of intrinsic variables. Not all perceptual control will, but if the organism "discovers" through random reconnection a perception that its actions can influence and therefore that it can control, it is more likely than not that there will be some occasions on which controlling that variable proves useful. So, I would argue, perceptual functions that produce controllable perceptions are more likely to survive future reorganization than are perceptual functions that produce signals that the organism has not found a way to control and that do not contribute to perceptions the organism has learned to control.
...
Perhaps asking questions about how to construct perceptual functions
is the wrong approach. We already know -- loosely -- how perceptual
functions are constructed. They are the result of random variations.
So isn't the construction of perceptual functions just a matter of
exploiting randomness in the right way?
I think so, if "exploiting" is interpreted appropriately.
"...but how it functions is not understood."
Do we really have to think about *how* a N-level system functions?
We may not "have to", but Bill did, and it is very much worthwhile to understand what he discovered and demonstrated, even if you then choose to follow a different path. What Bill showed (among much else) is that it is possible to create a multi-level multi-input multi-output control hierarchy with randomly weighted interconnections, and by random reorganization based on the ability to control, develop a system that effectively can bring the high-level perceptions under control using multiple input and outputs for each.
What happens if we setup the conditions for a control system that is
in chronic error until it starts to perceive some kind of N-level
variable?
I presume you mean "N-th level", not "N-level". There aren't any N-level variables, each perceptual variable being a scalar represented by the rate of firing in a specific nerve tract.
What defines a "level" in HPCT?
Follow the B:CP Chapter by Chapter seminar, or read B:CP by yourself. If you don't have a copy, I believe you can get one on-line. Simply put, in HPCT a level contains a type of perceptual function that is particular to that level and different from the type of perceptual function in other levels. The interconnections among control units are between consecutive levels and don't jump levels or connect within a level.
Can you have a system that overall seems to be "moving towards" a new
perceptual level?
Yes, if you drop "overall", which is just a condition in which more and more functions of a type are brought under control. Babies don't have all the levels that a mature person has. The levels develop as the person matures, but presumably by growing first one perceptual function at the new level and then more similar ones, not by creating all the perceptual functions of a level at once, instantaneously.
Martin
