| View previous topic :: View next topic |
| Author |
Message |
Guest
|
 Posted: Wed Nov 19, 2008 3:42 pm Post subject: training the network |
 |
|
Hi folks,
I have a very general question. What is the benefit actually to train
the neural network with genetic algorithm?I tried to train the neural
network with genetic algorithm for wind speed forecasting and wonder
what's the benefit of using genetic algorithm?
I will very grateful if someone will clear me out.
thanks.
ashraf |
|
| |
|
 Back to top |
Greg Heath
Guest
|
| Posted: Thu Nov 20, 2008 10:49 am Post subject: Re: training the network |
|
|
On Nov 19, 10:42 am, ashraf.l...@gmail.com wrote:
| Quote: | Hi folks,
I have a very general question. What is the benefit actually to train
the neural network with genetic algorithm?I tried to train the neural
network with genetic algorithm for wind speed forecasting and wonder
what's the benefit of using genetic algorithm?
I will very grateful if someone will clear me out.
|
Most training algorithms seek local minima. Therefore,
it is often necessary to make multiple weight initialization
runs to obtain satisfactory results.
Genetic algorithms are extremely slow structured searches
for global minima. I have not used them. However, I have
not seen demos that illustrate that the genetic algorithms
will yield timely results that are significantly superior
to the simple multiple weight initialization approach.
Hope this helps.
Greg |
|
| |
|
| Back to top |
Stephen Wolstenholme
Guest
|
| Posted: Thu Nov 20, 2008 6:04 pm Post subject: Re: training the network |
|
|
On Wed, 19 Nov 2008 07:42:28 -0800 (PST), ashraf.leaf@gmail.com wrote:
| Quote: | Hi folks,
I have a very general question. What is the benefit actually to train
the neural network with genetic algorithm?I tried to train the neural
network with genetic algorithm for wind speed forecasting and wonder
what's the benefit of using genetic algorithm?
I will very grateful if someone will clear me out.
thanks.
ashraf
|
In my experience using a genetic algorithm for neural network training
is a very slow process. Using a genetic algorithm for initialisation
may have some advantages but I have only tried it a few times. I have
not established if it has any general advantages.
Steve
--
Neural Planner Software Ltd http://www.NPSL1.com
JustNN. Just Neural Networks. http://www.justnn.com
EasyNN-plus. Neural Networks plus. http://www.easynn.com
SwingNN. Forecast with Neural Network. http://www.swingnn.com |
|
| |
|
| Back to top |
Greg Heath
Guest
|
| Posted: Sat Nov 22, 2008 9:12 pm Post subject: Re: training the network |
|
|
On Nov 20, 7:04 am, Stephen Wolstenholme <st...@tropheus.demon.co.uk>
wrote:
| Quote: | On Wed, 19 Nov 2008 07:42:28 -0800 (PST), ashraf.l...@gmail.com wrote:
Hi folks,
I have a very general question. What is the benefit actually to train
the neural network with genetic algorithm?I tried to train the neural
network with genetic algorithm for wind speed forecasting and wonder
what's the benefit of using genetic algorithm?
I will very grateful if someone will clear me out.
thanks.
ashraf
In my experience using a genetic algorithm for neural network training
is a very slow process. Using a genetic algorithm for initialisation
may have some advantages but I have only tried it a few times. I have
not established if it has any general advantages.
|
I agree ... especially if the topology is not simple
(e.g., multiple hidden layers, feedback, skip-layers,
fixed zero weights, etc). However, once the toplogy
and weights are initialized, the more traditional
optimization algorithms will be more efficient.
Hope this helps.
Greg |
|
| |
|
| Back to top |
|
