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

neural net prediction - pruning procedures

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

neural net prediction - pruning procedures

Postby numpsy » Fri Jan 30, 2009 7:06 pm


i've used your neural net prediction algo... i want to ask... are there any functions available to let the net decide which architectur is the best for it.

i read about the so called pruning procedures.

Posts: 3
Joined: Fri Jan 30, 2009 7:03 pm

Re: neural net prediction - pruning procedures

Postby andrew.kirillov » Sat Jan 31, 2009 1:41 pm


If we speak about AForge.NET framework, then there are no functions, which could suggest network size at this point. Choosing the right network size is not the simplest task. Some researchers made some suggestions and formulas, which could be used to calculate amount of layers and neurons, but still there is no ideal one. In many cases network size is chosen based on experiments, experience and some theoretical assumptions.

Network pruning is usually done after some initial network training. For example, a regular feed forward network may be taken (with connections all to all), initial training may be done, and then some connections between neurons may be removed. There are different purposes for this. One is to optimize network by removing connections, which don’t really affect network’s result. But the main purpose usually is to improve network’s ability to generalize data. This is also something to work on in the future versions …

With best regards,
User avatar
Site Admin, AForge.NET Developer
Posts: 3453
Joined: Fri Jan 23, 2009 9:12 am
Location: UK

Re: neural net prediction - pruning procedures

Postby kalori » Fri Feb 06, 2009 10:53 am

Also for reference see the NEAT (NeuroEvolution of Augmenting Topologies) project here: which evolves the architecture. There is an (apparently unmaintained) C# version that also implements pruning here: and can come up with surprisingly good results.

Posts: 4
Joined: Fri Feb 06, 2009 10:29 am

Return to Artificial Intelligence