ANN

My input array is

   35.0000    1.0000    1.2500    0.5000
   65.0000    1.0000    1.2500    0.5000
   35.0000    2.0000    1.2500    0.5000
   65.0000    2.0000    1.2500    0.5000
   35.0000    1.0000    1.7500    0.5000
   65.0000    1.0000    1.7500    0.5000
   35.0000    2.0000    1.7500    0.5000
   65.0000    2.0000    1.7500    0.5000
   35.0000    1.0000    1.2500    1.0000
   65.0000    1.0000    1.2500    1.0000
   35.0000    2.0000    1.2500    1.0000
   65.0000    2.0000    1.2500    1.0000
   35.0000    1.0000    1.7500    1.0000
   65.0000    1.0000    1.7500    1.0000
   35.0000    2.0000    1.7500    1.0000
   65.0000    2.0000    1.7500    1.0000
   20.0000    1.5000    1.5000    0.7500
   80.0000    1.5000    1.5000    0.7500
   50.0000    0.5000    1.5000    0.7500
   50.0000    2.5000    1.5000    0.7500
   50.0000    1.5000    1.0000    0.7500
   50.0000    1.5000    2.0000    0.7500
   50.0000    1.5000    1.5000    0.2500
   50.0000    1.5000    1.5000    1.2500
   50.0000    1.5000    1.5000    0.7500
   50.0000    1.5000    1.5000    0.7500
   50.0000    1.5000    1.5000    0.7500
   50.0000    1.5000    1.5000    0.7500
   50.0000    1.5000    1.5000    0.7500
   50.0000    1.5000    1.5000    0.7500

My output array is

    7.9700
   15.2200
    9.5300
   12.6000
    9.4400
   13.1600
    9.8100
   11.6500
    8.4700
   14.0500
    9.5000
   11.7000
    7.0900
   11.1000
    6.6900
   10.4400
    6.2900
   14.8800
   12.2500
   10.3800
   10.6500
    9.5000
   10.1600
   10.6500
   11.1300
   11.5000
   11.1000
   10.9900
   10.9800
   11.0300

I used Neural Network tool(nntool). I clicked new in the network data manager and in the create network page I selected the Network type as Feed-forward backprop

Input data in a variable'c' as given above and selected it from the dropdown box and the target data 'd' same. other parameters like training function & Adaption learning function and Performance function as default values.

If I need only one hidden layer, Do I need to set the number of layers as 1?

Similarly I set the number of neurons as 3 for layer1 & the transfer function as LOGSIG and created the network by name network1. Now I opened the network and clicked the Train tab. In the training info I selected the Training Data Input as c and target as d and clicked Train. But I am not getting the ANN predicted output value corresponding to each input value.

Please help me at the earliest.

Answer 1

[Malathy]

we use the SIM function to see the predicted response of the ANN.

[Walter Roberson]

husam eddin comments

me2

Answer 2

[Greg Heath]

If you plot output vs each of the inputs you will see that each plot consists of 3 vertical lines. The best model I can think of is a 2nd order polynomial that will intersect each line approximately in the middle.

Hope this helps.

Greg

[Malathy]

My question here is how to see the predicted response of ANN

Answer 3

[Greg Heath]

ANN : Malathy asked on 24 Feb 2012 at 7:49

My input array is

-----SNIP

My output array is

-----SNIP

I used Neural Network tool(nntool). I clicked new in the network data manager and in the create network page I selected the Network type as Feed-forward backprop

Input data in a variable'c' as given above and selected it from the dropdown box and the target data 'd' same. other parameters like training function & Adaption learning function and Performance function as default values.

If I need only one hidden layer, Do I need to set the number of layers as 1?

% NO. The number of weight layers is 2 (input-hidden and hidden- output)

Similarly I set the number of neurons as 3 for layer1 & the transfer function as LOGSIG

% TANSIG is preferred because inputs are transformed (via default MAPMINMAX) to be bipolar.

and created the network by name network1. Now I opened the network and clicked the Train tab. In the training info I selected the Training Data Input as c and target as d and clicked Train. But I am not getting the ANN predicted output value corresponding to each input value.

Please help me at the earliest.

% You'll have to quantify your performance.

% I ran 10 trials each for H = 0:4 (H=0 is just a linear regression but you need to use H = [] in newff). The summary statistics for an adjusted R^2 obtained from all of the data are tabulated below. If Ntrneq is the number of training equations and Nw is the number of weights to be estimated, the adjusted MSEtrn is obtained from SSEtrn/(Ntrneq-Nw)to reduce the optimistic bias caused by using the same data that was used in estimating the weights. Validation and Test error estimates are not usually adjusted even though the validation data is used in the design.

%Ignoring the default Ntrn/Nval/Ntst division 0.6/0.2/0.2, I just used all of the data to obtain MSEa = Neq*MSE/(Neq-Nw).

H = 0 1 2 3 4

minR2a = 0.7797 0.1428 0.1497 0.5269 0.4192

meanR2a = 0.8207 0.7342 0.7733 0.8689 0.8363

medianR2 = 0.8222 0.8199 0.8479 0.9062 0.9015

stdR2a = 0.0262 0.2267 0.2312 0.1461 0.1841

maxR2a = 0.8622 0.8599 0.9573 0.9969 0.9876

As you can see, for H = 3 and 4, the best of 10 designs yields R2a ~ 0.99.

Hope this helps.

Greg