how to decrease the neural network error

I have six input as natural frequencies while the output are the depth and location of damage in x and y direction . this means that the net has three outputs. when i trained the net the error is very high. how can I decrease the error. worth noting that I have 90 cases in total as input of the net.

x=Input; t=Target;

% Create a Fitting Network

net = feedforwardnet(10);

% Train the Network

net = train(net,x,t);

view(net)

output = net(x);

perf = perform(net,t,output)

Answer 1

[Greg Heath]

 close all, clear all, clc
 [ I N ] = size(input) % [ 6 90 ]
 [O N ]  = size(target)% [ 3 90 ]
 MSE00   = mean(var(target',1)) % Target Variance
 Ntrn    = N-2*round(0.15*N)    % 62 Training examples (default)
 Ntrneq  = Ntrn*O               % 186 Training equations
 H       = 10                   % number of hidden nodes (default)
 Nw      = (I+1)*H+(H+1)*O      % 103 Unknown weights
 r       = Ntrneq/Nw            % 1.8 May want to consider reducing H
 MSEgoal = 0.01*MSE00           % net will model 99% of target variance

Both the initial weights and the data division are random. You probably have obtained an unfortunate choice of random initial weights. Therefore you should design 10 or more nets for H=10 and choose the best ones according to the validation set error tr.best_vperf. Where

 [ net tr ] = train(net,input,target);

If you are going to apply this net to new data, you may want to consider reducing H to increase r and make the net more robust.

I have posted many, many examples of double for loop designs over number of hidden nodes h = Hmin:dH:Hmax and number of random weight initializations i=1:Ntrials.

For regression and curve-fitting search the NEWSGROUP (and ANSWERS) using

 greg fitnet Ntrials

For classification and regression search using

 greg patternnet Ntrials

Hope this helps.

Thank you for formally accepting my answer

Greg