/*

* This code demonstrates the example by Matt Mazur:

* https://mattmazur.com/2015/03/17/a-step-by-step-backpropagation-example/

* Code by Stuart Cording

* aka codinghead@gmail.com

*/

Neural network;

void setup() {

// This neural network uses two inputs, two hidden nodes, and two output nodes.

network = new Neural(2,2,2);

// Set learning rate here

network.setLearningRate(0.5);

// Set Neural class to be 'verbose' so we can see what it is doing

network.turnVerboseOn();

println("Matt Mazur Neural Network Backpropagation Example");

println("-------------------------------------------------");

println();

println("Structure is:");

println("Input nodes: ", network.getNoOfInputNodes(), "; Hidden nodes: ", network.getNoOfHiddenNodes(),

"; Output nodes: ", network.getNoOfOutputNodes());

// Set network biasing here

// b1 = 0.35

network.setBiasInputToHidden(0.35);

// b2 = 0.60

network.setBiasHiddenToOutput(0.6);

// The Neural class constructor gives the weights random value.

// Here we use the values from Matt Mazur's example.

// We start with the input-to-hidden weights:

// w1 = 0.15 (i1 to h1)

network.setInputToHiddenWeight(0, 0, 0.15);

// w3 = 0.25 (i1 to h2)

network.setInputToHiddenWeight(0, 1, 0.25);

// w2 = 0.20 (i2 to h1)

network.setInputToHiddenWeight(1, 0, 0.2);

// w4 = 0.30 (i2 to h2)

network.setInputToHiddenWeight(1, 1, 0.30);

// Next we configure the hidden-to-output weights:

// w5 = 0.40 (h1 to o1)

network.setHiddenToOutputWeight(0, 0, 0.4);

// w7 = 0.50 (h1 to o2)

network.setHiddenToOutputWeight(0, 1, 0.5);

// w6 = 0.45 (h2 to o1)

network.setHiddenToOutputWeight(1, 0, 0.45);

// w8 = 0.55 (h2 to o2)

network.setHiddenToOutputWeight(1, 1, 0.55);

// Configure the inputs

// i1 = 0.05

network.setInputNode(0, 0.05);

// i2 = 0.10

network.setInputNode(1, 0.1);

// Now declare the values we like to achieve at the outputs for this

// input combination

// o1 should be 0.01

network.setOutputNodeDesired(0, 0.01);

// o2 should be 0.99

network.setOutputNodeDesired(1, 0.99);

// We now perform a forwardpass using the configured inputs, weights

// and bias value. Verbose is on, so you will see the working

println();

println("Calculating values for o1 and o2...");

println();

network.calculateOutput();

println();

// Let's summarise the current state

println("...forwardpass complete. Results:");

println("For i1 = ", network.getInputNode(0), " and i2 = ", network.getInputNode(1));

println("o1 = ", network.getOutputNode(0), " (but we want: ", network.getOutputNodeDesired(0), ")");

println("o2 = ", network.getOutputNode(1), " (but we want: ", network.getOutputNodeDesired(1), ")");

println();

println("Total network error is: ", network.getTotalNetworkError());

println();

// Now we'll perform a learning cycle using backpropagation

// This enables a backpropagation cycle everytime the calculateOuput() method is called

network.turnLearningOn();

println("Learning is ON");

println("Calculating outputs again and performing backpropagation...");

println();

network.calculateOutput();

// We'll turn learning off again...

network.turnLearningOff();

// ...and run another forwardpass without verbose on:

network.turnVerboseOff();

network.calculateOutput();

println("...backpropagation complete. Results:");

println("For i1 = ", network.getInputNode(0), " and i2 = ", network.getInputNode(1));

println("o1 = ", network.getOutputNode(0), " (but we want: ", network.getOutputNodeDesired(0), ")");

println("o2 = ", network.getOutputNode(1), " (but we want: ", network.getOutputNodeDesired(1), ")");

println();

println("Total network error is: ", network.getTotalNetworkError());

println();

}

void draw() {

}