Neural-Networks

We perform analysis of the sonar data using neural networks. For 100 independent replications (i.e., 100 independent training /testing splits),

• Fit four different neural networks:

-Two distinct single hidden layer neural networks: (i) A single hidden layer with 15 nodes (ii) A single hidden layer with 18 nodes:

-Two distinct neural networks with two hidden layers: (iii) Two hidden layers with 23 and 20 nodes and (iv) Two hidden layers with 28 and 23 nodes.

• Compare the accuracy of these four Neural networks among them.

• Also compare it to other classification methods. Specifically, we will compare neural network performance to the methods below:

  Linear Discriminant Analysis (LDA). Quadratic Discriminant Analysis (QDA). Regularized Discriminant Analysis(RDA).

  Support Vector Machines (SVMs):

  Naïve Bayes.

Data

 Sonar Dataset. To generate in R Code, we use : library(mlbench) and data(Sonar)

Code implementation.

R packages: neuralnet 

• The comparison will be possible by implementing simulated accuracy matrices. We calculate the average accuracy for each method and compare these values.

Conlusion

-Among the four Neural Networks, the one with the best performance is the Neural Network with Two hidden layers with 28 and 23 nodes (NeuralNet_28_23).

-Among all the classification methods The accuracy of the Superlearner method is the best. Followed by RandomForest and ( Neural Network with Two hidden layers with 28 and 23 nodes) NeuralNet_28_23. Then, NeuralNet_18 , NeuralNet_15, Gradient Boosting method etc.