MiniProj3
RBF Classification
You will use radial basis function neural networks as classifiers for
east cancer detection.
The dataset is as follows:
First, read the dataset description at:
http:
archive.ics.uci.edu/ml/datasets/Breast+Cancer+Wisconsin+%28Diagnostic%29
For your convenience, I have reformatted the data for the MATLAB. The input matrix is called P,
each row of which co
esponds to different measurements of a patient’s tumor cell sample. The
co
esponding element in the output vector T is -1 if the cell was determined to be benign, and
+1 if malignant.
Load P and T into MATLAB’s by using the menu “File > Import Data” option, or by simply
dragging and dropping.
Creating RBF neural nets: load P and T into MATLAB, and divide it into training, validation,
and test subsets (60%, 20%, and 20% ratios) using the following command
[trainP,valP,testP,trainInd,valInd,testInd] = dividerand(P,0.6,0.2,0.2);
[trainT,valT,testT] = divideind(T,trainInd,valInd,testInd);
Now create two RBF neural nets, exact and regular, using
net1 = new
e(trainP,trainT,SPREAD1);
net2 = new
(trainP,trainT,GOAL,SPREAD2);
Please read MATLAB’s help documentation for new
e and new
for more
information. You can find the output and MSE e
or for each network by (e.g.
validation, network 1) y1v = sim(net1,valP);
mse1v=mse(y1v-valT);
Or for network 2 training:
y2tr= sim(net2,trainP);
mse2tr=mse(y2tr-trainT)
Note that you have to choose a SPREAD (and GOAL, for new
) first.
Do this: change the values of SPREAD (and GOAL, for new
) until you get a validation MSE
of 0.75. Report on your observations. Specify the network (
or
e, SPREAD, size, and GOAL)
that got you to that point. Report the test MSE of your final solution. Plot the training, validation,
and test ROC curves for the above network.
Note: you may want to start with larger spreads, e.g. 10 or more, and use a loop for next values.
Deliverable : ALL the requested results (numbers, ROC curves, observations); as well as
your conclusions on each section. Also, compress and submit ALL your saved neural
network objects and programs (but not the training data).
P.mat
P:[30x569 double a
ay]
T.mat
T:[1x569 double a
ay]