neural network poor performance

Hello this is the first time I work with the neural network toolbox, I designed a network using newff, the goal is to approximate the input vector wich is a 4x600 matrix of MAV's taken from 4 muscles, to an output an expected angle.

I followed the instructions given at: http://www.mathworks.com/matlabcentral/answers/137-how-do-i-improve-my-neural-network-performance

however when I look at the regression plot, I'm getting a very low regression index, R=0.16882, the fit line it's almost horizontal and there is a lot of dispersion.

If someone could point me out in the right direction, I'd be gratefull.

Here is my code:

net = newff(Input,Target,20);

net.trainParam.goal=1e-6;

net.trainParam.max_fail=6;

net.performFcn='msereg';

net.performParam.ratio=0.5; 

net,tr] = train(net,Input,Target);

yTargets = sim(net,Input);

plotregression(Target,yTargets);

[noclah]

Also, I'm using trainlm, an I noticed that it stops, because it reaches validation checks to quickly.

[Hessam]

I think you should change the parameter"net.divideFcn"to "dividerand". It's the default option for the nn when you do not specify the divide function explicitely. then you should use the following parameters to change the ratios of each part:

net.divideParam.trainRatio = 50/100;

net.divideParam.valRatio = 5/100;

net.divideParam.testRatio = 35/100;

However the validation check(the following parameter() is a very tricky part:

net.trainParam.max_fail

Hessam

Answer 1

[Greg Heath]

Find, H, the smallest number of hidden nodes that will yield a satifactory design by looping over nH candidate values and trying Ntrials different designs for each value of H. Then choose the best of the nH*Ntrials designs.

I typically look at nH = 10, Ntrials = 10 first. Then accept or modify depending on the result.

Hope this helps.

Greg

[noclah]

Thanks Greg, I have tried that too. I choose to change the network structure. Now I have a 4-20-20-1 network, this time something else happened, when I use the plotregression command, I get a very good regression index for the training set .9998 after 500 epochs, however for the validation and test set, I get .007 and .19 respectively, I suspect I need to modify the error, but I still dont know how to do it.

Here is my network:

net =

Neural Network object:

architecture:

numInputs: 1

numLayers: 3

biasConnect: [1; 1; 1]

inputConnect: [1; 0; 0]

layerConnect: [0 0 0; 1 0 0; 0 1 0]

outputConnect: [0 0 1]

numOutputs: 1 (read-only)

numInputDelays: 0 (read-only)

numLayerDelays: 0 (read-only)

subobject structures:

inputs: {1x1 cell} of inputs

layers: {3x1 cell} of layers

outputs: {1x3 cell} containing 1 output

biases: {3x1 cell} containing 3 biases

inputWeights: {3x1 cell} containing 1 input weight

layerWeights: {3x3 cell} containing 2 layer weights

functions:

adaptFcn: 'trains'

divideFcn: 'dividerand'

gradientFcn: 'gdefaults'

initFcn: 'initlay'

performFcn: 'mse'

plotFcns: {'plotperform','plottrainstate','plotregression'}

trainFcn: 'trainlm'

parameters:

adaptParam: .passes

divideParam: .trainRatio, .valRatio, .testRatio

gradientParam: (none)

initParam: (none)

performParam: (none)

trainParam: .show, .showWindow, .showCommandLine, .epochs,

.time, .goal, .max_fail, .mem_reduc,

.min_grad, .mu, .mu_dec, .mu_inc,

.mu_max

weight and bias values:

IW: {3x1 cell} containing 1 input weight matrix

LW: {3x3 cell} containing 2 layer weight matrices

b: {3x1 cell} containing 3 bias vectors

other:

name: ''

userdata: (user information)

Train Parameters:

show: 50

showWindow: 1

showCommandLine: 0

epochs: 1000

time: Inf

goal: 0.0100

max_fail: 500

mem_reduc: 1

min_grad: 1.0000e-010

mu: 0.0500

mu_dec: 0.1000

mu_inc: 10

mu_max: 1.0000e+010

I'm getting close, yet I'm still too far. Thanks.

[Greg Heath]

>Thanks Greg, I have tried that too.

Hmm. Since that has ALWAYS worked for me, I would like to see your code.

Hope this helps.

Greg

Answer 2

[Mo al]

you have to show your code. very quick suggestions 1- increase your performance goal. 2- initiate your network

Answer 3

[Hessam]

Just as a very important point in my work: How does the Performance function calculate the "mse"? is it normalized? otherwise it would be useless.

[Greg Heath]

MSE = mse(target-output)

It can be normalized by MSE00, the MSE obtained from the naive constant model:

output00 = repmat(mean(target,2),1,Ntrn)

MSE00 = mse(target-output00)

NMSE = MSE/MSE00 % normalized MSE

R2 = 1 - NMSE % R^2 statistic

Useless might be an appropriate characterization for multiple output

nets when the outputs have different scales. Therefore, if outputs have different scales, it would be wise to normalize them.

For regression with multiple real-valued outputs, I prefer standardization (zero-mean/unit-varance).

For classification among c classes, using targets that are columns of the c-dimensional unit matrix eye(c) is sufficient.

Hope this helps.

Greg