From: Eugene Leitl (Eugene.Leitl@lrz.uni-muenchen.de)
Date: Wed Mar 28 2001 - 02:22:14 MST
On Tue, 27 Mar 2001, James Rogers wrote:
> The "lookup table", as it is being used here, is a perfectly optimal and
> correct model, but useless because it requires mind-boggling quantities of
> memory. (Actually, "lookup table" has been poorly defined thus far; I
Right, since you can't just map sensorics to motorics without using hidden
state (gobs and oodlefuls of hidden state, which also acts as part of the
input, making it a rather nondeterministic lookup table, if you're just
looking at world-visible I/O stream), or using a current-age light cone as
vector input, which is a tad wasteful.
> don't think everyone is using it the same way.) Writing heuristics to
> replace sections of the table is a sub-optimal solution (provably so) and
> would have serious consequences if implemented on a broad scale.
Actually, you can prove that that table is anything but random, unless
compressed optimally. There also smart lookup tables, which use all
manners of interpolation.
This archive was generated by hypermail 2.1.5 : Sat Nov 02 2002 - 08:06:44 MST