using System;
using System.Collections.Generic;
using MathNet.Numerics;
using MathNet.Numerics.LinearAlgebra;

namespace NNet
{
	class Program
	{
		public static void Main (string[] args)
		{
			var input = new List<Vector<double>> () {
				Vector<double>.Build.DenseOfArray(new double[] {-1.0, -1.0}),
				Vector<double>.Build.DenseOfArray(new double[] {1.0, 1.0})
			};
			var output = new List<Vector<double>> () {
				Vector<double>.Build.DenseOfArray(new double[] {-1.0}),
				Vector<double>.Build.DenseOfArray(new double[] {1.0})
			};

			NeuralNetwork net = new NeuralNetwork (2, 1);
			for (int epoch = 0; epoch < 10; epoch++) {
				for (int i = 0; i < input.Count; i++) {
					var result = net.Feed (input[i]);
					net.Train (output [i].Subtract (result));
					Console.WriteLine (result);
				}
				Console.ReadLine ();
			}
		}
	}

	public class NeuralNetwork
	{
		Vector<double> input;
		Vector<double> output;
		Matrix<double> weight;
		Matrix<double> rmsprp;

		Func<double, double> Activate = Math.Tanh;
		Func<double, double> dActivate = Differentiate.DerivativeFunc (Math.Tanh, 1);

		public NeuralNetwork (int inputs, int outputs)
		{
			weight = Matrix<double>.Build.Random (rows: outputs, columns: inputs);
			rmsprp = Matrix<double>.Build.Random (rows: outputs, columns: inputs).Map(Math.Abs);
		}

		// input = -1.0 to 1.0 range
		public Vector<double> Feed (Vector<double> input)
		{
			this.input = input;
			output = weight.Multiply (input);
			return output.Map (Activate);
		}

		// error = expected - output
		public Vector<double> Train(Vector<double> error)
		{
			Vector<double> deriv = error.PointwiseMultiply (output.Map (dActivate));
			Matrix<double> delta = deriv.OuterProduct (input);
			delta.PointwiseMultiply (delta).Multiply (0.05).Add(rmsprp.Multiply(0.95),rmsprp);
			weight.Add (delta.PointwiseDivide(rmsprp.Map(Math.Sqrt)).Multiply(0.1), weight);
			return weight.LeftMultiply (deriv);
		}
	}
}