help in training a nueral network

hello; i need to train a feed forward neural network using trainlm ;i used this code but i did not get the result until i used 650 neurons in the hidden layer which is too big , i need some one to help me to reduce this number and get the results

p=[ 28220       22142       28395       23145];
t=[24930       25444       25958       26472];
net=newff(minmax(p),[100,1],{'tansig','purelin'},'trainlm');
%[net,TR] = trainlm(net,Pd,Tl,Ai,Q,TS,VV,TV)
%info = trainlm(code)
net.trainParam.epochs=100
net.trainParam.goal=0.01
net.trainParam.max_fail=1000
net.trainParam.mem_reduc=100
net.trainParam.min_grad=10e-10
net.trainParam.mu=0.0001
net.trainParam.mu_dec=0.001
net.trainParam.mu_inc=100
net.trainParam.mu_max=100e10
net.trainParam.show=25
net.trainParam.time=inf
[net,tr]=train(net,p,t)
y2 = sim(net,p)

thnx

Answer 1

[Greg Heath]

[ I N ] = size(p) % [1 4]

[ O N ] = size(t) % [ 1 4]

Neq = N*O % 4 No. of training equations

Nw = (I+1)*H + (H+1)*O % No of unknown weights for I-H-O node topology.

When training to convergence without regularization (e.g., TRAINBR), and/or Early Stopping with a validation set, require Neq >= Nw or H <= (Neq-O)/(I+O+1) = 1

However, desire Neq >> Nw or H << (Neq-O)/(I+O+1) to mitigate noise and neasurement errors.

Therefore, either get more data, use H = 1 or use a linear model (H=0).

Only override the net = newff(..) and net.trainParam.* defaults if absolutely necessary. See doc newff and doc train or doc trainlm for a list of defaults.

Hope this helps.

Greg.

P.S. Try fitting one period of a sine wave.