Mathematics / Mathematik / Matemática

• About this group
• Forums

• Home
• Groups
• Mathematics / Mathematik / Matemática
• Forums
• Analysis
• Neural network for data interpolation

• 
  Hendrik Gieseke

  The company name is only visible to registered members.

  Neural network for data interpolation

  I'm currently working on a electromagnetic field analysis tool (wireless lan) to display and interpolate samples captured with kismet wireless.
  
  The aim of this application is to drive through large company areas, search for accesspoints, locate them and generate a graphic report for the IT security.
  
  The current development state is:
  * import of kismet files is working
  * sorting and classification of data is working
  * analysis based on the mathematical inverse distance weighted calcaulation works
  * graphical output works
  
  Application details:
  Language: C# (.Net 3.5)
  Database: DB4O (filebased object database)
  XML-Handling: Altova source from dtd-files
  Neural Network: NeuronDotNet / Encog
  License: GPL, LGPL, or anything else (without charge)
  
  As mathematical analysis is only correct for free space, I currently try to implement the analysis based on a neural network (feedforward backpropagation)
  
  As the samples have high error rates, I think, I'll have to implement a SOM network, to cluster the data and remove points with high errors.
  
  Is there anybody with knowlage, how to implement this, as my current implementaion does not seem to work at all.
  
  Input-Data for the network for learning:
  Longitude as double, Latitude as double, Signal as byte casted to double.
  
  I tried a 2-x-x-1 network, where I used from 5 to 50 neurons for x. The network was not able to learn correctly.
  I tried with norming the inputs and outputs (range -1 to 1 and -0.5 to 0.5), but this didn't solve the problem.
  The problem might be the high error rate, but I still got no working solution for a SOM network, to remove high error points.
  
  I'm searching for somebody with experience in neural network implementations for data interpolation.
  
  Best regards
  Hendrik
  
  • 27 Feb 2009, 08:56 am
  Hendrik Gieseke wrote: > I'm currently working on a electromagnetic field analysis tool > (wireless lan) to display and interpolate samples captured with > kismet wireless. > > The aim of this application is to drive through large company areas, > search for accesspoints, locate them and generate a graphic report > for the IT security. > > The current development state is: > * import of kismet files is working > * sorting and classification of data is working > * analysis based on the mathematical inverse distance weighted > calcaulation works > * graphical output works > > Application details: > Language: C# (.Net 3.5) > Database: DB4O (filebased object database) > XML-Handling: Altova source from dtd-files > Neural Network: NeuronDotNet / Encog > License: GPL, LGPL, or anything else (without charge) > > As mathematical analysis is only correct for free space, I currently > try to implement the analysis based on a neural network (feedforward > backpropagation) > > As the samples have high error rates, I think, I'll have to implement > a SOM network, to cluster the data and remove points with high errors. > > Is there anybody with knowlage, how to implement this, as my current > implementaion does not seem to work at all. > > Input-Data for the network for learning: > Longitude as double, Latitude as double, Signal as byte casted to > double. > > I tried a 2-x-x-1 network, where I used from 5 to 50 neurons for x. > The network was not able to learn correctly. > I tried with norming the inputs and outputs (range -1 to 1 and -0.5 > to 0.5), but this didn't solve the problem. > The problem might be the high error rate, but I still got no working > solution for a SOM network, to remove high error points. > > I'm searching for somebody with experience in neural network > implementations for data interpolation. > > Best regards > Hendrik
• 
  Dr. Marcel Blattner    Premium Member

  The company name is only visible to registered members.

  Re: Neural network for data interpolation

  Perhaps you can try other methods for clustering, like spectral graph clustering with a normalized laplacian. The method
  works quite fine in other problem settings.
  Cheers
  Marcel
  
  • 28 Feb 2009, 08:39 am
  Dr. Marcel Blattner wrote: > Perhaps you can try other methods for clustering, like spectral graph > clustering with a normalized laplacian. The method > works quite fine in other problem settings. > Cheers > Marcel
• 
  Dr. Axel Kemper

  The company name is only visible to registered members.

  Re: Neural network for data interpolation

  They following paper (cf. page 7) could be helpful:
  http://www.sensorsportal.com/HTML/DIGEST/march_08/Special_Is...
  
  Why do you insist on a neural network?
  Wouldn't it make sense to design a model for the possible error parameters and minimize the expected error according to your model? Looking back at Gauss and his least squares method to reduce measurement errors, I would advocate for some clever scheme of oversampling to construct your maximum likelihood estimator.
  
  Another article in this direction:
  http://ieeexplore.ieee.org/Xplore/login.jsp?url=/iel5/8/1764...
  
  Good luck!
  
  Axel Kemper
  
  • 28 Feb 2009, 2:50 pm
  Dr. Axel Kemper wrote: > They following paper (cf. page 7) could be helpful: > http://www.sensorsportal.com/HTML/DIGEST/march_08/Special_Issue_Vol_90/P_SI_38.pdf > > Why do you insist on a neural network? > Wouldn't it make sense to design a model for the possible error > parameters and minimize the expected error according to your model? > Looking back at Gauss and his least squares method to reduce > measurement errors, I would advocate for some clever scheme of > oversampling to construct your maximum likelihood estimator. > > Another article in this direction: > http://ieeexplore.ieee.org/Xplore/login.jsp?url=/iel5/8/17643/00814957.pdf?arnumber=814957 > > Good luck! > > Axel Kemper
• 
  Hendrik Gieseke

  The company name is only visible to registered members.

  Re^2: Neural network for data interpolation

  First to say - thanks for the ideas.
  I'll have a look at the provided links and will try to get further. But as this is a spare time project, it will take a few days :-)
  
  • 28 Feb 2009, 9:14 pm
  Hendrik Gieseke wrote: > First to say - thanks for the ideas. > I'll have a look at the provided links and will try to get further. > But as this is a spare time project, it will take a few days :-)