|
I want to perform wave height prediction/forecasting with artificial neural networks (ANNs). The inputs should be the present & previous wave heights and the past & previous wind speeds. I would like to use MATLAB. In the case where this is a problem I would consider to use python as it appears to have a larger selection of libraries to various ANN algorithms. I would like to have results using conjugate gradient (cgd), and cascade correlation. The neural net toolbox of matlab appears to have cgd, but not cascade correlation. I found this post with code for cascade correlation HERE, so could I simply use this? Is there another implementation I could use because I cannot find it. I have only seen cascade correlation in neural networks applied with python link which refers to a package. Would convolutional neural networks be applicable too link to package description? They seem to be inspired by vision problems where most of the applications are done. How is the regulation of the internal complexity of the ANN controlled? Are there some parameters for AIC which can be used or some other technique? |
The advantage of CNNs for vision problems is that they exploit locality and translation equivariance: the fact that nearby pixels in images are typically more correlated, and the fact that filters learnt in one part of the image will also likely be useful elsewhere in the image. Of course I don't know what your data looks like exactly, but it's not readily apparent to me if it has these properties, and how they would manifest themselves if so.
The only way I see for a convolutional approach to work in this context is when you have multiple "time series" of wave lengths and corresponding wind speeds - then a convolution in the time dimension might make sense. This approach was used by Honglak Lee et. al for audio classification (audio signals are also time series).