Differrence between feed forward & feed forward back propagation

I used neural netowrk MLP type to pridect solar irradiance, in my code i used fitnet() commands (feed forward)to creat a neural network.But some people use a newff() commands (feed forward back propagation) to creat their neural network. please what's difference between two types?? :

 net=fitnet(Nubmer of nodes in haidden layer); --> it's a feed forward ?? true??   

 net=newff(Nubmer of nodes in haidden layer); ---> it's a feed forward back propagation ??

Help me please wchich one can i choose for my case (prediction problem)???!!! Who appripriate??

Best regards

[Greg Heath]

A USEFUL POST

GREG

[YERRUPALLI GANESH]

Difference between levenberg matquardt and feed forward backpropagation algorithm

Answer 1

[Greg Heath]

1. Regardless of how it is trained, the signals in a feedforward network flow in one direction: from input, through successive hidden layers, to the output.

2. Given a trained feedforward network, it is IMPOSSIBLE to tell how it was trained (e.g., genetic, backpropagation or trial and error)

3. A feedforward backpropagation net is a net that just happened to be trained with a backpropagation training algorithm. The backpropagation training algorithm subtracts the training output from the target (desired answer) to obtain the error signal. It then goes BACK to adjust the weights and biases in the input and hidden layers to reduce the error.

4. Current examples of feedforward nets are

 a. FITNET for regression and curve-fitting which calls the generic FEEDFORWARDNET
 b. PATTERNNET for classification and pattern-recognition which calls the generic FEEDFORWARDNET
 c. FEEDFORWARDNET

5. OBSOLETE examples of feedforward nets are

 a. NEWFIT for regression and curve-fitting which calls the generic NEWFF
 b. NEWPR for classification and pattern-recognition which calls the generic NEWFF 
 c. NEWFF

6. The default training functions for the above algorithms use backpropagation. However, the designer can train the nets any way they want.

7. The Neural Network Toolbox does not offer a genetic algorithm (GA). However, there are posts to the NEWSGROUP and ANSWERS regarding training with a GA

SEARCH WORDS NEWSGROUP HITS ANSWERS HITS

NEURAL GA 69 39

NEURAL GENETIC 132 56

8. None of the above are recommended for timeseries forecasting. The CURRENT timeseries forecasting algorithms are

 a. TIMEDELAYNET: Feedforward net with output that only depends on current and past inputs
 b. NARNET:       Feedback net with output that only depends on past outputs
 c. NARXNET:      Feedback net with output that depends on BOTH past outputs as well as current and past inputs

Hope this helps.

Thank you for formally accepting my answer

Greg

P.S. Online documentation for any function can be obtained using the commands help, doc and type. For example

 help fitnet
 doc fitnet
 type fitnet

[omar belhaj]

 Thank you very much Greg

Please Gerg if i understand these are the things that I have learned

 1- feedforward backpropagation it's mean i used backpropagation algoritm to train my network (true or false idea).

 2- feedforward it's mean i used any algorithm to train the network

3- For me i will use which one fitnet(inputs,targets)or feedfowrednet ()?? can i obtain the same results ???

[Greg Heath]

1. My name is Greg

2. I do not recommend using backpropagation in the name of a net. Given a net you cannot tell what kind of algorithm was used to train it.

3. Feedforward means there are no paths where signals travel backwards or sideways.

4. Use fitnet and patternnet. Then there is no question of the type (regression/curve-fitting or classification/pattern-recognition)

Hope this helps.

Greg

[omar belhaj]

So please i will just to know fitnet() and feedforwardnet() are the similar commands to create the neueral networks?? because i found a lot of researches use feedforwardnet() instead fitnet()for prediction issues !!!!

[Greg Heath]

Although both FITNET and PATTERNNET call FEEDFORWARDNET, each have different defaults w.r.t. their specific function, i.e., regression/curve-fitting vs classification/pattern-recognition.

So, just use FITNET and PATTERNNET.

Answer 2

[khan]

what you said it suggest that the new function have both forward and backward propagation in the same function. Where is the old have only forward pass. You should right click and select help on each of them and you will see.

[omar belhaj]

I ask whcich one is appropraite to use for forecasting ?? whcich the best ??

[Greg Heath]

@Khan

No. the new functions are just revised versions of the old ones. If you use the commands DOC, HELP, and/or TYPE, you will see that NEWFIT, NEWPR and NEWFF are OBSOLETE and as such, will no longer be maintained by MATLAB.

@omar:

For forecasting/predicting the future, use TIMEDELAYNET(only inputs), NARNET(only output feedback) or NARXNET(both inputs and output feedback). Again, use DOC, HELP, and/or TYPE for details.

Greg

Answer 3

[Shireen Shah]

can you please tell me which command(newff or feedforwardnet) is good choice to be used for channel estimation in OFDM using neural networks?

[Greg Heath]

Use

 FITNET for regression

and

 PATTERNNET for classification

Both are special cases of FEEDFORWARDNET.

Corresponding NEWFIT, NEWPR and NEWFF are obsolete and have been replaced by the above.