how to solve this problem using neural network?

Hi I'm pretty new to Matlab but it seems easy to understand Matlab language and I've some skills in programming like Basic.

first I need to create a simple neural network program that can be trained, so it can get some input of:

[age(years), weight(kg), height(cm)] and decide how much this person fit to be an athlete from 1 to 10. for example:

[20, 70, 180] = 10 % OK

[85, 95, 170] = 1 % Not OK

where should I begin from? how much input, hidden and output neuron do I need?

Answer 1

[Greg Heath]

 For N pairs of I-dimensional "I"nputs and corresponding 
O-dimensional "O"utput target outputs, the matrices must 
have the dimensions

 [ I N ] = size(input)
 [ O N ] = size(target) % = size(output)

Then

 Ntrn    = N -2*round(0.15*N)   % default No of training vectors
 Ntrneq  = Ntrn*O               % default No of training equations

For an I-H-O network the number of unknown weights is

 Nw = (I+1)*H + (H+1)*O = O + (I+O+1)*H

Therefore, the number of training equations is greater or equal to the number of unknowns when H, the number of hidden nodes satisfies

 Ntrneq >= Nw   ==> H <= Hub 

 where the hidden node upper bound, Hub, is given by 

 Hub = (Ntrn-1)*O/(I+O+1)

When H > Hub, the net is OVERFIT and there is a danger of OVERTRAINING the overfit net and losing the ability to generalize performance beyond the training samples.

The best ways to avoid overtraining an overfit net are

 1. Do not overfit by using H <= Hub
 2. THE MATLAB DEFAULT: 
    a. H = 10
    b. Use VALIDATION STOPPING, which stops training when 
the validation subset error increases for more than 6 
(The MATLAB default ) continuous epochs. 
 3. BAYESIAN REGULARIZATION: Use the training function 
TRAINBR which uses a penalized training goal 
SSE + const * sum(unknown weights) to penalize the number 
and size of unknown weights.

Your problem will probably be best solved as a classification problem using PATTERNNET with a one dimensional target that is either 0 or 1.

1. See the documentation

 help patternnet
and
 doc patternnet

2. See some of my examples by searching BOTH the NEWSGROUP and ANSWERS with

 greg patternnet

Hope this helps.

Thank you for formerly accepting my answer

Greg

[H.misaq]

I really Need a more simple answer, as I said I'm "pretty new to Matlab". could you please be so kind and provide me some simple code example based on my input stracture that I've mentioned? I heard that I need to go trought multilayer perceptron thing... Thanks for your time.

[Greg Heath]

There is nothing simpler than to read and try the examples in the help and doc documentation.

 help patternnet

 doc patternnet

Next try searching in both the NEWSGROUP and ANSWERS for posted questions and answers. The magic search word is

 PATTERNNET

however, YOU CAN REDUCE THE NUMBER OF POSTS BY USING

 GREG PATTERNNET

Hope this helps,

Greg