A simple way to guess the number of hidden neurons comes from mathematics. As you know we needs at least same numbers of equation and variables to solve a multi-unknown variable equation set. In neural networks weights are your unknown variables and each training data make an equation then the sum of all the weights in the neural network at least must be equal to your input training data.

For example if you have 3 input neurons, 6 hidden neurons and 2 output neurons you have 40 (3x6 + 6x2) unknown variable that must be estimated by your network. In this example at least you need 40 training data. However, in practice you need two to three times of 40 training data to reach an appropriate result. This approach won't give you the optimal number of neurons immediately but at least you have a good starting point.I advise to start with 36 input training data (in this example) and increase the number of training data gradually. In each step check the system for over fit.After a few try you will find the optimal answer.

In your case, if I get it correctly, you have 14 input neurons and one output neurons. It seems that in best situation you can use 4 hidden neuron. I think 4 hidden neuron is not enough and you have to find more training data. Usualy, SVM works better than MLP in case of low traning dada.

Dear Masrur,

I do not know about any paper that suggest minimum number of data for ANN. But based on my experience at least 20 data is required for modelling. You can check some of my papers. If you have 80 data, in my opinion that is enough for ANN modelling.

I hope this answer would be helpful for you.

A simple way to guess the number of hidden neurons comes from mathematics. As you know we needs at least same numbers of equation and variables to solve a multi-unknown variable equation set. In neural networks weights are your unknown variables and each training data make an equation then the sum of all the weights in the neural network at least must be equal to your input training data.

For example if you have 3 input neurons, 6 hidden neurons and 2 output neurons you have 40 (3x6 + 6x2) unknown variable that must be estimated by your network. In this example at least you need 40 training data. However, in practice you need two to three times of 40 training data to reach an appropriate result. This approach won't give you the optimal number of neurons immediately but at least you have a good starting point.I advise to start with 36 input training data (in this example) and increase the number of training data gradually. In each step check the system for over fit.After a few try you will find the optimal answer.

In your case, if I get it correctly, you have 14 input neurons and one output neurons. It seems that in best situation you can use 4 hidden neuron. I think 4 hidden neuron is not enough and you have to find more training data. Usualy, SVM works better than MLP in case of low traning dada.

There is no specific rules for training the ANN model. If you have more number of observation we have higher degree of accuracy in prediction or otherwise less.

Dear A. A. Masrur Ahmed,

This question is very interesting.

In my opinion, the suitable answer depends on the real system and many experimental factors. For instance, in a multilayer perceptron or similar supervised ones, you have to take into account the ANN topology and transfer function used (among other things). I mean, regarding the topology, the number of observations used in the training phase has to be higher than the number of weights. Regarding the training function, there are training functions that preventing malfunctions better than others. To sum up, these thing have to be solved by the experience (experimental designs, testing different number of observations, etc.)

I hope it has helped you.

Cheers,

José

Dear A. A. Masrur Ahmed,

In my experience, the richer the dataset, the better the ANN performance. I would say that having only 80 samples for 14 variables is a bit on the thin side. Would you have the possibility to collect more experimental data?  It also depends on how strong each variable's effect is on the output and also on how much noise there is in the data. For noisy datasets , your training set needs to be much larger.

You may find these books useful: 

Understanding Neural Networks

by John Iovine

Code Your Own Neural Network: A step-by-step explanation

by Steven C. Shaffer

Good luck!

Luiza

The following link may be useful for you.

http://www.heatonresearch.com/node/707

Hi...i agree with iman's answer...adding that if you have 14 parameters then input layer has 14 neurons and one output neuron as per the problem statement. Even a small network with 3 hidden neurons will have 50 weights or free parameters. Having only 80 data points to estimate 50 parameters is highly unjustified mathematically. Summarising, you should not attempt this problem with ANN...

hey , 

I invite you to use my MATLAB code

hope it will help you

Article MATLAB Code of Artificial Neural Networks Estimation