There are different types of network creation functions which can be employed in artificial neural network such as cascade-forward network, competitive neural layer, distributed delay network, elman network, feed-forward network, function fitting network, layered recurrent network, linear neural layer, learning vector quantization (LVQ) network, nonlinear auto-associative time-series network, radial basis network and self-organizing map.
One of the most common networks is feed-forward backpropagation neural network. Is it better than the other networks in any aspect? If so, what are the advantages? And does this network have any noticeable drawback?
Feed forward neural networks process signals in a one-way direction and have no inherent temporal dynamics. Thus, they are often described as being static.
Dear Mr Iranpour, The basic definition will be as stated by Mr Islam. The other main type is Recurrent Networks in which there is flow of information or feedback from output to input. Feed forward networks are relatively simpler to implement. These have numerous applications where nonlinear mapping is to be done between inputs and outputs and predicting future states e.g. nonlinear behavioral modeling, adaptive control, image recognition etc. You cannot use feed forward ANN where the outputs are also dependent upon previous state of inputs as well which are also referred to as memory effects. If you want to use ANN also familiarize yourself with ANN training methods. You can easily find basic tutorials on ANN on the internet. If you want to design one I will recommend go through the documentation of Matlab Neural Network Toolbox. In case of any specific questions I will be willing to help further. Regards
Dear Maximilian, dear George, dear Md. Mohaimenul and dear Irfan,
Thank you all for your answers.
I intend to use artificial neural network to derive empirical equations which correlate inputs to output. For this purpose, I used to employ feed-forward backpropagation network trained by Levenberg–Marquardt algorithm. As far as I know, using this network can only employ LogSigmoid or TanSigmoid transfer functions in the hidden layer which leads to complicated forms of equations, especially in the case of using more nodes in the hidden layer.
Is there a better alternative to derive equations which leads to equations with simpler forms and acceptable level of accuracy?
These days people use piece-wise linear transfer functions instead of sigmoidal. And put more hidden layers. On the other hand, it is always better to have some theory (even semi-empirical) to connect inputs with outputs.
Dear Mr Iranpour, The way I understand your problem I will recommend that you try using a single layer MLP. Train it in Matlab. Once you meet your desired level of convergence note the value of the weights and biases by reading the cell structure. Now the equation for the FF ANN (including biases) will form your empirical equations. There is no problem if you use Tansigmoid transfer function on hidden layer and linear function on the output layer. So your empirical equations will feature these functions. You can even use the exponential form equivalent of tansigmoid function. My personal experience is you can get good results for 80 -90% problem using a single layer percepton with sufficient number of neurons. You can use multiple layers that will just extend the size of equations. Regards
- Badges
- Science topic