RE: Parallelizing attention and credit-assignment

From: Ben Goertzel (
Date: Fri Sep 09 2005 - 15:45:15 MDT

> Ben - It isn't well-known to me. Can you tell me a reference to
> read so I can understand why you're saying this? My impression
> is that the average engineer would say that
> reinforcement learning or temporal-differernce learning are
> the best means of credit assignment, and the average AI researcher
> would say the bucket brigade or a truth-maintenance and belief-
> revision system are the best means, all of which are parallel
> approaches.
> That systems implementing the bucket brigade aren't intelligent
> doesn't mean that the bucket brigade doesn't work.
> - Phil

Some of the fundamental conceptual flaws with the bucket brigade algorithm
are well dealt with in Eric Baum's work, see his papers on
artificial economies of agents at

I don't really remember which ideas are contained in which papers,

"Manifesto for an Evolutionary Economics of Intelligence"

is probably a good start.

And also I suppose he mentions this material
in "What is Thought?" though
he doesn't give full technical detail there.

Unfortunately, however, Baum's solutions to the problems he points
out with the bucket brigade don't really work well either. They
do solve problems the bucket brigade can't, as he demonstrates via
his Blocks World examples, but they fall apart on larger examples
(perhaps due to lack of scalability and tunability rather than for
fundamental mathematical reasons ... but even so...).

One of the Novamente team reimplemented Baum's Hayek system for
reinforcement learning and replicated some of Baum's results but
not others, and found the system extremely slow and very sensitive
to its parameters....

So far as I know Baum is not putting too much of his energy into
the ongoing development of Hayek, which would seem to indicate he
has also recognized its severe limitations, though he hasn't
explicitly acknowledged them in print.

-- Ben

This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:52 MDT