Re: Is the Curt net a kind of decision tree?


Curt Welch wrote:
"JGCASEY" <jgkjcasey@xxxxxxxxxxxx> wrote:
< ... >


So how is it different from a decision (or classification) tree?

Well, it isn't different. It IS a decision tree (I've said this
multiple times) - except all decision trees I know have only one
entry point and since these nets have multiple entry points they
are really not one decision tree but multiple decisions tress
overlapping. And the branches of most decisions trees don't
recombine where this happens all over these nets. So it's an odd
sort of decision tree but that's basically what it is.

And most important, it's a learning system so there are also
algorithms at work which constantly change the behavior of the
decision tree over time. Most decision trees are static things
which don't learn.

Looking at these networks as decisions trees trained by reinforcement
learning would be a valid way to look at them. The ultimate problem
that all intelligent machines must solve is the answer to the question
"what should I do now"? Using huge decision trees to answer that
question seems reasonable to me. That's basically what these networks
are.


Yes I understand the "logic" behind your net which to me is a witches
brew where you have mixed in some Dr Who temporal magic ideas, mixed
together some decision trees, mixed in some stm "timestamps" and some
ltm threshold values, smoothed the mixture out with some ratio values,
at the same time saying you shouldn't think in terms of "values" but
only "temporal events" and then go on talking about values and then
top it all off with some magical incantations which are your
explanations of the "power" of the net.


As part of the creative process all that is acceptable but now is
the time for some rigorous analysis.


But, if you have read my other post, you realize I've finally seen
what seems to be an important weakness in the design of these nodes.
They are blind to the classifications that would require knowing the
source of each pulse. they can't react differently to an AB sequence
than to a BA sequence. And I'm now thinking this is important to be
able to do. So I have to look at this more to see if the tests these
nets are doing are lacking some important powers - and to see if I can
find some simple way to give them that power that still allows for
simple reinforcement training of the decision.


As far as I know you don't have any "simple" reinforcement apart from
the worst kind of hill climbing algorithms for you net?


Although your nodes, or neurons for that matter, don't "know" and
cannot "know" where the pulses come from, this can still be given
a place code by the system. The "knowledge" is not in the nodes but
can be in the connections.


So when you say, an AB sequence, you are referring to the fact that
pulse A and pulse B have a *place code* not just a temporal existence.
This information is destroyed by your nodes when a pulse passes
through them. The pulse that comes out of the node provides no
information as to where it came from. The information as to it
being event A or event B is destroyed as that is not the kind of
information your nodes extract. Neurons are able to extract this
information with lateral connections, such as the requirement for
determining motion and its direction.


I think you need to think in terms of what kind of information the net
actually does extract from the input (stimuli) in order to make its
"decisions" (responses).


In my "frequency sorter" the outputs provide information as to
the size range of the gaps entering the input. Allow that to merge
and that information is destroyed which is the effect I believe
your merging does most of the time. In the case of the "frequency
sorter" the resolution is reduced by the size of the net just as
an image resolution is reduced by the number of pixels making up
the image.


The kind of analysis I would have liked someone like Michael Olea
to have done was what kind of information processing a curt-like
net could do rather then the limitations of your current topology.


It also surprised me that Michael Olea was surprised that, even in
a random curtnetron, structures would emerge. Just look at the ever
changing patterns that form on a window pane when it rains. Or
the patterns carved out by streams coming down from the mountains.
The source of these patterns is in the rules, not in the random
inputs to the net. Where do the patterns of Conway's game of Life
come from? Not the from the input for it doesn't have an input.
My nets also display inherent patterns resulting from the rules
governing their internal interactions.


However I think your notion is of a universal behavior generator
that can be honed to produce only intelligent behaviors via some
external critic without the existence of special purpose modules.


I think curtnetron is a better name as curtron sounds like a name
of a nuclear particle. You don't suffer from physics envy do you?


--
JC

.