|
PureBytes Links
Trading Reference Links
|
I worked at Sprint for a few years and we used genetic algorithms to
optimize phone networks. I wrote a demonstration program that we used to
demostrate to our customers. It solved the travelling salesman problem
using a GA. For comparison, I also wrote a brute-force, try every solution
program. The brute-force algorithm took about 30 minutes to find the
shortest route between 9 random points on a 2D grid and it came up with the
correct answer 100% of the time (of course, because it tried every possible
solution). The GA program took 30 seconds to find it's answer and it found
the best solution on the same set of data about 80% of the time. In those
cases where it didn't find the shortest route, it always found a route that
was pretty close to shortest. That's one caveat with GA's: they don't
always find the best solution, but they usually get very close if you run
them correctly. Short of matrix reduction, they are the most powerful
search/optimization technique I've ever seen and they are surprisingly
simple in theory.
One of the guys from Ward Systems hangs out on the comp.ai.neural-nets
newsgroup lurking for posts like yours. :) Frankly, I don't understand his
rationale for using a GA to optimize the NN input data set, but I haven't
seen his software so there might be a good reason for it. I recently built
a program that trains standard back-prop NN's using a GA and so far I
managed to prove that this technique will successfully curve-fit the nnet to
the data. :) At any rate, his responses in the newsgroup seem well
reasoned and he's happy to answer questions, but that doesn't really mean
anything about his software. My guess is that it is probably like anything
else for sale for trading: garbage-in, garbage-out. On their website, they
are displaying trades made from a system generated by their system
generator; sure looks impressive. And another thing that adds to
credibility is I believe they are applying the same technology to other
diverse problems like supply-chain management and pumping systems. But
again with trading software, I think 98% of the problem is knowing how to
feed it.
Kent
-----Original Message-----
From: Guy Tann <grt@xxxxxxxxxxxx>
To: Metastock <metastock@xxxxxxxxxxxxx>
Date: Thursday, January 27, 2000 12:47 AM
Subject: Ward Systems' NeuroShell Trader Professional
List,
Has anybody had any experience with this software?
My brother found this article about genetic algorithm testing. They use it
for optimizing where you need to run a bunch of variables and in regular
simulations where the computer would run for months. Sort of like the human
genome project.
Thanks,
Guy
|