MATLAB: Create a PDF file that includes your analysis and answers to the questions posed below concerning the given neural network class using the iris data. You will be asked to create and train networks of a few different configurations. Try changing the maxepochs $($training time$)$ and learnrate $($learning rate$)$ values to investigate how quickly and well the networks may be trained. Ideally, you want the network to train quickly $($in few epochs$)$ and not have a "spikey" RMS plot $($adjust the learning rate to see if you can smooth it out$).$ The network class will create plots and return all the necessary data. Include the training and classification plots as part of your discussion.

How to Use the Class

Network Creation:

n $=$ neuralnet$([43],$ 'Transfer', 'sigmoid', 'data', 'iris.mat'$)$

Network Training:

n$.$Train$($maxepochs$,$ learnrate, true$)\%$ true creates the plots for training and classification

Network Testing:

$[$rms$,$ cr$]=$ n$.$Test$($test$,$ true$)\%$ rms is the root mean square error, cr is classification rate, true gives case$-$by$-$case data

This command may also be used by replacing 'test' with 'train' to see the information about the training set. The plots generated while training use the 'train' values from this function.

Iris Classification Visualization:

n$.$PlotIris $\%$ This just shows the data, it won$$t be affected by the network structure or training

Configurations to Try

Single$-$Layer Network

Create a single$-$layer network with $4$ inputs, $3$ outputs, a sigmoid transfer function, and the iris data. What happens when you try to train the network in this configuration? Does it solve part of the problem or fail completely? Make sure to check the case$-$by$-$case information.

Multi$-$Layer Networks

Try the following: $[433],[443],[453],[4333],$ and $[4433].$ Does one of these work better than the others $($that is$,$ have a higher classification rate or train faster to a low error$)?$ Try different amounts of training and learning rate values to see if some train in less time. Typical choices for maxepochs are between $500$ and $5000$ and learnrate are between $0\mathrm{.}001$ to $0\mathrm{.}1.$

ReLU Activation

Choose your best$-$performing $2-$ and $3-$layer network from above and try them with a ReLU activation by replacing $$sigmoid$$ with $$relu$$ when creating the network. Compare the training and accuracy with the sigmoid version and discuss your findings. Make sure to adjust the training time and learning rates since ReLU will have different optimal values.

There are $8$ Network Configurations to Report on:

$[43]$ sigmoid

$[433]$ sigmoid

$[443]$ sigmoid

$[453]$ sigmoid

$[4333]$ sigmoid

$[4433]$ sigmoid

The best of $[433],[443],[453]$ with ReLU

The best of $[4333],[4433]$ with ReLU