Introducing Jeff

Introducing Jeff Hunter.

  Hello CGS folks. I'm a contractor working within earshot of
Martin. Naturally I have been evangelized about (and converted to) PCT.

  My PCT credentials:
  I've been a lurker on the CSG net for many months. I've also
read Powers' Byte articles, but not any of his books. Martin (Taylor) ,
Allan (Randall), and I have hashed over PCT a fair bit over coffee.
I've also looked over the shoulder of Chris Love while he has built
the Little Baby (the learning Little Man). I hand-coded a set of ECS's
(Elementary Control Systems) that let the Baby put finger to target
(purely for testing the software, not to cheat on the learning side).

  My PCT motivation:
  I am most interested in PCT as a method of partitioning the
learning of medium sized nets (100 nodes) on non-trivial tasks.
To explain: the size of neural nets is hitting a (soft) ceiling. If you
double the number of nodes (to handle a more complex task) you more
than double the training time. The current wisdom on solving this
problem is to divide the net in half, train each half of the net on half
the task, and then add some more net "on top" that takes two half-answers
and makes one whole. Naturally this only works in some domains, but
it seems to be a win.
  The big drawback is that the designer of the net must cut the
problem domain in half "by hand".

  PCT seems to be a natural fit for this technique. We start with
some low-level control systems and train them on simple tasks. Then
we add a higher level control system and train it to use the lower
levels on a more complicated task. Repeat ad nauseum.

  If we can get re-organization figured out we may not even have
to partition the task by hand. (This is a bit optimistic. When teaching
human children we usually start them simple and work up.)

  My projected PCT postings:
  I have a few new contributions (of varying profundity) to the
ongoing debate. I also have a lot of backlogged followups to postings.

  I've done a few quick-and-dirty programs in my spare time.
I may get to spend a while concentrating on PCT models at work.

        ... Jeff

···

--
De apibus semper dubitandum est - Winni Ille Pu