private static void TestPlotting()
{
ScatterBuilder.BuildAndDumpScatters(new List <ScatterInfo>()
{
new ScatterInfo("Jorge title", "Y AXIS NAME", "X AXIS NAME", new List <Series>()
{
new Series(new List <double>()
{
1, 2, 3, 4, 5, 6, 7, 8
}, "jorge series 1"),
new Series(new List <double>()
{
8, 7, 6, 5, 4, 3, 2, 1
}, "jorge series 2"),
})
});
}
static void Main(string[] args)
{
// TestPlotting();
var randomCategoriesSeparableProblem = new RandomNumberCategoriesSeparable();
var randomCategoriesInseparableProblem = new RandomNumberCategoriesInseparable();
var perceptronProblemPairs = new List <PerceptronProblemPair>()
{
// SimplePerceptron, different learning rates
new PerceptronProblemPair(new SimplePerceptron(new StepFunction(0), 100, 0.01), randomCategoriesSeparableProblem),
new PerceptronProblemPair(new SimplePerceptron(new StepFunction(0), 100, 0.0001), randomCategoriesSeparableProblem),
new PerceptronProblemPair(new SimplePerceptron(new StepFunction(0), 100, 0.00005), randomCategoriesSeparableProblem),
// Pocket is better than no pocket
new PerceptronProblemPair(new PocketSimplePerceptron(new StepFunction(0), 100, 0.0001), randomCategoriesSeparableProblem),
new PerceptronProblemPair(new SimplePerceptron(new StepFunction(0), 100, 0.0001), randomCategoriesSeparableProblem),
// SimplePerceptron vs. Pocket in non-separable. Overfitting on SimplePerceptron, but not on Pocket
new PerceptronProblemPair(new PocketSimplePerceptron(new StepFunction(0), 1000, 0.0001), randomCategoriesInseparableProblem),
new PerceptronProblemPair(new SimplePerceptron(new StepFunction(0), 1000, 0.0001), randomCategoriesInseparableProblem),
// Thermal and simple - Inseparable, thermal better
new PerceptronProblemPair(new ThermalSimplePerceptron(new StepFunction(0), 100, 0.0001, 10), new RandomNumberCategoriesInseparable()),
new PerceptronProblemPair(new SimplePerceptron(new StepFunction(0), 100, 0.0001), new RandomNumberCategoriesInseparable()),
// Multilayer fitting a few points - demonstrate capabilities of multilayer perceptron
new PerceptronProblemPair(new MultiLayeredPerceptron(0.5, 1000, 2), new CoordinateFitting()),
new PerceptronProblemPair(new MultiLayeredPerceptron(0.5, 1000, 4), new CoordinateFitting()),
new PerceptronProblemPair(new MultiLayeredPerceptron(0.5, 1000, 8), new CoordinateFitting()),
new PerceptronProblemPair(new MultiLayeredPerceptron(0.5, 1000, 16), new CoordinateFitting()),
new PerceptronProblemPair(new MultiLayeredPerceptron(0.5, 1000, 32), new CoordinateFitting()),
// Multilayer fitting a few points. Works very nice.
new PerceptronProblemPair(new MultiLayeredPerceptron(0.5, 2000, 2), new CoordinateFitting()),
new PerceptronProblemPair(new MultiLayeredPerceptron(0.5, 2000, 4), new CoordinateFitting()),
new PerceptronProblemPair(new MultiLayeredPerceptron(0.5, 2000, 8), new CoordinateFitting()),
new PerceptronProblemPair(new MultiLayeredPerceptron(0.5, 2000, 16), new CoordinateFitting()),
new PerceptronProblemPair(new MultiLayeredPerceptron(0.5, 2000, 32), new CoordinateFitting()),
// SimplePerceptron and NAND
new PerceptronProblemPair(new SimplePerceptron(new StepFunction(0), 50, 0.03), new NandLogicalGate()),
new PerceptronProblemPair(new PocketSimplePerceptron(new StepFunction(0), 50, 0.03), new NandLogicalGate()),
new PerceptronProblemPair(new MultiLayeredPerceptron(1, 1000, 4), new NandLogicalGate()),
// PocketPerceptron is better than simpleperceptron at XOR, but none can converge on it
new PerceptronProblemPair(new SimplePerceptron(new StepFunction(0), 1000, 0.05), new XorLogicalGate()),
new PerceptronProblemPair(new PocketSimplePerceptron(new StepFunction(0), 1000, 0.05), new XorLogicalGate()),
// XOR gate and how it behaves against Multilayer
new PerceptronProblemPair(new MultiLayeredPerceptron(0.5, 3000, 1), new XorLogicalGate()),
new PerceptronProblemPair(new MultiLayeredPerceptron(0.5, 3000, 2), new XorLogicalGate()),
new PerceptronProblemPair(new MultiLayeredPerceptron(0.5, 3000, 4), new XorLogicalGate()),
new PerceptronProblemPair(new MultiLayeredPerceptron(0.5, 3000, 8), new XorLogicalGate()),
new PerceptronProblemPair(new MultiLayeredPerceptron(0.5, 3000, 16), new XorLogicalGate()),
new PerceptronProblemPair(new MultiLayeredPerceptron(0.5, 3000, 32), new XorLogicalGate()),
// Overfitting when trying to make all neurons work in inseparable data
new PerceptronProblemPair(new MultiLayeredPerceptron(1.5, 2000, 1), randomCategoriesInseparableProblem),
new PerceptronProblemPair(new MultiLayeredPerceptron(1.5, 2000, 2), randomCategoriesInseparableProblem),
new PerceptronProblemPair(new MultiLayeredPerceptron(1.5, 2000, 4), randomCategoriesInseparableProblem),
new PerceptronProblemPair(new MultiLayeredPerceptron(1.5, 2000, 8), randomCategoriesInseparableProblem),
new PerceptronProblemPair(new MultiLayeredPerceptron(1.5, 2000, 16), randomCategoriesInseparableProblem)
};
// NOTE: Can be parallelized easily via .AsParallel()
List <ScatterInfo> perceptronResults = new List <ScatterInfo>();
foreach (PerceptronProblemPair perceptronProblemPair in perceptronProblemPairs)
{
IList <ScatterInfo> result = perceptronProblemPair.Perceptron.SolveProblem(perceptronProblemPair.Problem);
perceptronResults.AddRange(result);
}
ScatterBuilder.BuildAndDumpScatters(perceptronResults);
}