georgekosmidis / simpleneuralnetwork

A C# multi-layer perceptron (one input, multiple hidden and output), forward feed neural network with backward propagation and x number of neurons for each layer.

Simple Neural Network in C#

AI

A multi-layer perceptron (one input, one output and multiple hidden layers), forward feed neural network with backward propagation and x number of neurons for each layer.

Characteristics

• 1 Input, [1-..] Hidden, 1 Output layer
• Unlimited neurons per layer
• Forward feeding, backward propagation
• 2 Methods, HyperTan / Sigmoid
• Set number of hidden layer and neurons or allow automatic adjustment to find the optimal solution
• Samples are divided in Train data, Verification data and Test data
• Save / Load model from disk
• Custom Modeler with instuctions (SimpleNeuralNetwork.ProblemModeler/Problems/Custom.cs)
• One model with 3 input neurons and 2 output neurons (SimpleNeuralNetwork.ProblemModeler/Problems/AddSubtract.cs)
  • Output Neuron 1 substracts the three input values,
  • Output Neuron 2 adds them
• One model with 49 input neurons and 49 output neurons for Lotto number predictions (SimpleNeuralNetwork.ProblemModeler/Problems/Lotto.cs)
  • Output Neurons 1-49, show probability for each number

Math Methods

Two available methods depending on the model.

• Sigmoid as output method is in in the range of 0 to 1, so input/ouput data must me normalized from 0 to 1
• HyperTan is in in the range of -1 to 1, so input/ouput data must me normalized from -1 to 1

Program.cs

Working example of how to train the Neural Network to add and substract three decimals

How to

var neuralNetwork = ConsoleHelper.TrainAndReturnNetwork<[IProblemLotto | IProblemAddSubstract | IProblemCustom]>(bool SaveTrainedNetwork);
// - OR -
var neuralNetwork = ConsoleHelper.LoadAndReturnNetwork<[IProblemLotto | IProblemAddSubstract | IProblemCustom]>();

Test NN efficiency by trying unknown numbers as variables with Run:

//Example for IProblemAddSubstract
var result = new NeuralNetworkRunnerFactory()
	.Get()
	.Run(neuralNetwork, new double[ 7, 8, 3 ]);

Use Custom Model in SimpleNeuralNetwork.ProblemModeler/Custom.cs to model your own problem.

//Values of Output neurons define the expected result of the neural network
//Read the values of the model vertically to have the functions: 
//                                                               f( 2, 1, 1 ) = [  0, 4 ] 
//                                                               f( 3, 2, 1 ) = [  0, 6 ]
//                                                               f( 2, 1, 2 ) = [ -1, 5 ]
//                                                               f( 1, 1, 1 ) = [ -1, 3 ]
//                                     Values for Input Neuron 1----^   
//                                     Values for Input Neuron 2------^
//                                     Values for Input Neuron 3---------^
//                                    Values for Output Neuron 1------------------^
//                                    Values for Output Neuron 2--------------------^
//For example, Input neuron 1 will have as input '2', neuron 2 will have '1', and neuron 3 will have '1'
//Expected value for Output neuron 1 is '0' and for Output Neuron 2 is '4'
//Neural Network will try to replicate procedure f for every unknown input. That's what NN do :)
return new ProblemDescriptionCreator()
	//-----------------------------------------------------------------------------------
	//.SetMathFunctions(MathFunctions.HyperTan)         //Set the algorithms to be used                         
	//.SetHiddenNeurons(5)                              //Set the number of hidden neurons
	//--OR--                                
	.AutoAdjustHiddenLayer()                            //Let the network handle hidden neurons in order to find optimal solution
	//-----------------------------------------------------------------------------------

	.SetAcceptedError(.02)                              //Set accepted error for the train session to complete, current is 1%
	.SetNeuralNetworkName("Custom")                     //Set Network Name

	.AddInputNeuron(x => x.AddValues(2, 3, 2, 1))       //Add Input Neuron 1
	.AddInputNeuron(x => x.AddValues(1, 2, 1, 1))       //Add Input Neuron 2
	.AddInputNeuron(x => x.AddValues(1, 1, 2, 1))       //Add an Input Neuron 3

	.AddOutputNeuron(x => x.AddValues(0, 0, -1, -1))    //Add Output Neuron 1
	.AddOutputNeuron(x => x.AddValues(4, 6, 5, 3))      //Add Output Neuron 2

	.Get();                                             //Get the model

Output after training