How do i fix my neuronal network

Hello everybody, this is my first post so i hope im not breaking any rule or something like that.

Situation: Im working on a neuronal network on matlab (Not using nntool) and after 3 months i cant make it work as i expected.

Details: Is a classification Network with 45 inputs (all numeric) and one output (0, 0.5 and 1)

Problem: i have create my network and it works but im getting 63% of good responses and ofcourse this is not acceptable for my work. I have tried adding multiples hidden neurons (from 1 to 25) and the response is always de same, even if i change my inputs from 45 to 22 or 2. Also i tried to add epochs and change my activation function. but it didnt work aswell. Im really desperate since i have working on this for while and i cant get a solution.

My net: net = newff([inputs], [10 1] ,{'purelin','poslin'}); Inputs are numeric and there are no negative values.

Example of input: 1 2 6 2 1 2 3 0 31 6 16 0 0 0 414 411 250 9 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 2 0 0 0 0 0 0 0

i dont know if you need more details but feel free to ask.

Any kind of help would be appreciate.

Thanks you in advance to everyone.

PS: I need to use neuronal network thats the only requeriment for my project.

[Adam]

Just for future reference:

https://uk.mathworks.com/matlabcentral/answers/29922-why-your-question-is-not-urgent-or-an-emergency

Posts that don't tell us your work is urgent are far more likely to be reacted to more quickly.

Answer 1

[Greg Heath]

1. Since training functions transform inputs into the interval [-1 1], the hidden node functions should be odd, AND non-linear. TANH or TANSIG (THEY ARE THE SAME FUNCTION) fills the bill.

2. For classifiers the targets sould be considered probability densities conditional on the input

3. For classifiers of c distinct classes, the targets should be columns of the c-dimensional unit matrix eye(c). The corresponding output function is SOFTMAX.

4. If classes are not distinct, each target is in the interval [ 0 1 ] and the sum of the targets are unity ( e.g. [ 0.1 0.3 0.4 0.2 ]' ). The corresponding output function is LOGSIG.

5. The current classification training function is PATTERNNET which is based on FEEDFORWARDNET.

6. The previous classification function was NEWPR which is based on NEWFF.

Hope this helps.

Thank you for formally accepting my answer

Greg

PS: I have posted ~500 examples in the NEWSGROUP and ~50 in ANSWERS

[Ricardo Gonzalez]

Thanks you for the answer, im trying to apply those changes but it looks like there are some differences in the net because when i use Tansig on hidden and softmax on outputs im having this instant network training because minimum gradient has been reach

[Greg Heath]

Quite a lot depends on the random data division and random weight initialization. Therefore you should explicitly set the RNG at the top of a design loop and design multiple nets that only depend on the initial state of the RNG.

I typically design 10 nets for each value for the number of hidden nodes. The only difference is the initial state of the RNG.

search NEWSGROUP and ANSWERS for

 greg patternnet Hmin:dH:Hmax

Hope this helps.

Greg