AForge.NET

  :: AForge.NET Framework :: Articles :: Forums ::

a question about ElasticNetworkLearning

The forum is to discuss topics from different artificial intelligence areas, like neural networks, genetic algorithms, machine learning, etc.

a question about ElasticNetworkLearning

Postby whitebird » Tue Dec 21, 2010 9:11 am

Hi andrew!
i have tested the TSP sample and got an excellent result, then ,i change network learning arithmetic from ElasticNetworkLearning to SOMLearning, the result is really bad, all path crossed each other.
i have read the code of ElasticNetworkLearning class and SOMLearning class in AForge.Neuro.Learning namespace. one of main diffierences of this two learning arithmetics is the definition of distance, or more exactly,the difinition of coordinates of every neuron.
according to the difinition in ElasticNetworkLearning, all neurons form a ellipse.
in this definition of coordinates, after Learning, line every neuron orderly, why there is no any cross? can you say some principle of ElasticNetworkLearning?
with best wishes!
whitebird.
whitebird
 
Posts: 7
Joined: Wed Dec 15, 2010 3:56 pm

Re: a question about ElasticNetworkLearning

Postby andrew.kirillov » Tue Dec 21, 2010 10:03 am

Hello,

SOMLearning represents classical Kohonen SOM, which is a 2D map of neurons forming a grid. ElasticNetworkLearning represents ring (ellipse) looking network. When you train SOM, then on every learning step network updates not only winning neuron, but also some neighbor from the grid. However, when you train elastic network it updates also neighbors from the left/right only - there is no grid structure. So it kind of changes the original ring's shape, but it still stays a connected "ring", which shows the Traveling Sales Man’s route.

Here is some info, which may be useful.
With best regards,
Andrew


Interested in supporting AForge.NET Framework?
User avatar
andrew.kirillov
Site Admin, AForge.NET Developer
 
Posts: 3434
Joined: Fri Jan 23, 2009 9:12 am
Location: UK

Re: a question about ElasticNetworkLearning

Postby whitebird » Tue Dec 21, 2010 12:16 pm

thanks very much!
whitebird
 
Posts: 7
Joined: Wed Dec 15, 2010 3:56 pm




Return to Artificial Intelligence