For the past few months I've been working on the design of a recurrent
neural network for use in signaling formation changes in groups of
robots. The project is focused on machine learning, but I've naively tied
it to robotics so it appears to be related to my research :) What I
specifically want to do is generate neural networks with a genetic
algorithm, and evaluate their performance by looking directly at the
structure of the neural network rather than its output values.
what do you mean by "formation changes?"
Imagine a hallway that grows more narrow as you walk down it. Now assume
that you have a group of robots that can travel in, and change, formations.
The goal is to traverse the hallway with as many robots abreast as
possible. I want to use a neural network as a signalling mechanism for
changing formations.
dude, that is an awful lot of solution for this problem...
why would you want to evaluate the structure of the network rather than
the output values, isn't the whole point to use the GA to generate some
structure that's too complicated/non-obvious to have preconveived that
outputs the desired solution? what's the point of having a fancy network
that doesn't produce anything?
I want to analyze the structure of the network because I'd like a code base
that is capable of that. It could be used later for pruning networks,
producing statistical info about the structure, and maybe some other
stuff... *heheh*.
hrm... is this a vague possible future idea or something that you've already
given thought to? what would you use as a fitness test? ok so you want a
high-level neural network compiler and general tools to inspect them...
Your GA comment is correct.
The network itself may not produce anything, but the network analysis
software will. What I'm analyzing the networks for is a packet of activity
in the neurons - a state. I could then map this state to something useful.
Unfortunately, the runtime of the analysis tool will probably be pretty
bad... so I'm not sure what good that will do.
so start small...
what tools are you using for your GP stuff? lisp? :-) are you reading
koza's books? :-)
Not Lisp, although I'm intereted in persuing it. I'm using the Pygene
library... I know it's slow... I'll look at mosquito and CL over the break.
http://www.freenet.org.nz/python/pygene/
I've not read Koza's books yet, but I have watched one of his videos on
genetic programming. It was... dry... and... uhhh... it was a freaking
movie on genetic programming! Cool stuff though.
pygene looks cool...
I asked a friend of mine who is into both GP and Lisp, he said the language
of Mosquito Lisp is far too domain specific for security and networking to
be useful for something like GP and the only way to go is CL and Koza's
books (but of course he would say that, cl is a black hole people don't like
leave ;-)
I'm integrating multi-hop routing algorithms with 802.11a/b/g equipment
to produce wireless meshes. I'm then then building a robust
communication mechanism based on the publish/subscribe paradigm on top of
the meshes. It's a big project, and will take quite a while to complete.
is this an exercise in design or implimentation?
Implementation. None of the technology is new. In fact, most of the
algorithms have already cropped up in robotics. Unfortunately, there
aren't codebases to pull from for the application layer protocols, so I'll
have to roll all of it myself.
cool, well good luck
Nick
hrm, well, cool... back to homework I guess...
Damned homework! I can't wait for it to stop!
- Sebastian
_________________________________________________________________
MSN Shopping has everything on your holiday list. Get expert picks by style,
age, and price. Try it!
http://shopping.msn.com/content/shp/?ctId=8000,ptnrid=176,ptnrdata=200601&tcode=wlmtagline
_______________________________________________
RLUG mailing list
RLUG@rlug.org
http://lists.rlug.org/mailman/listinfo/rlug
