diff --git a/Reinforcement_Learning/NN_with_Backtracking.py b/Reinforcement_Learning/NN_with_Backtracking.py new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/Reinforcement_Learning/Perceptrons.py b/Reinforcement_Learning/Perceptrons.py index 67671654e43ac57b4bbb3078682644d849de13d3..7028ba4caf03b6ed02c6d0f762e36e6bcb1c4420 100644 --- a/Reinforcement_Learning/Perceptrons.py +++ b/Reinforcement_Learning/Perceptrons.py @@ -1,8 +1,9 @@ import numpy as np import Training_data +import matplotlib.pyplot as plt rng = np.random.default_rng(123) -TEACHDATA = 9999 +TEACHDATA = 99999 TESTDATA = 999 T_NUMBER = 6 # Number to be detected 0-6 @@ -31,6 +32,14 @@ class Neuron: self.input_count = input_count self.activation = activation self.weights = rng.random(input_count + 1) + self.errors = [] + + def plot_errors(self, titel): + plt.plot(self.errors) + plt.xlabel('Iteration') + plt.ylabel('Error') + plt.title(titel) + plt.show() def test(self, data): ix = np.insert(data.ravel(), 0, 1) @@ -49,12 +58,13 @@ class ThresholdPerceptron(Neuron): ix = np.insert(teach_data[j][i].ravel(), 0, 1) RI = self.activation(ix.dot(self.weights)) if RI != T: - delta = ETA * \ - (T - self.activation(ix.dot(self.weights))) * ix + err = T - self.activation(ix.dot(self.weights)) + delta = ETA * err * ix self.weights = self.weights + delta - if np.linalg.norm(old_weights - self.weights) == 0.00: - return self.weights - return self.weights + self.errors.append(abs(err)) + # if np.linalg.norm(old_weights - self.weights) == 0.00: + # return + return class SGDPerceptron(Neuron): @@ -65,16 +75,17 @@ class SGDPerceptron(Neuron): for i in range(TEACHDATA): old_weights = np.copy(self.weights) delta = [0 for _ in range(len(old_weights))] - for j in rng.choice(rng.permutation(len(teach_data)), 3): + for j in rng.choice(rng.permutation(len(teach_data)), 5): T = (j == T_NUMBER) ix = np.insert(teach_data[j][i].ravel(), 0, 1) - z = ix.dot(self.weights) - RI = self.activation(z) - delta = ETA * (T - RI) * RI * (1 - RI) * ix + RI = self.activation(ix.dot(self.weights)) + err = T - RI + delta = ETA * err * RI * (1 - RI) * ix + self.errors.append(abs(err)) self.weights += delta - # if np.linalg.norm(old_weights - self.weights) == 0.00: - # return self.weights - return self.weights + if np.linalg.norm(old_weights - self.weights) == 0.00: + return + return class LinearPerceptron(Neuron): @@ -88,8 +99,10 @@ class LinearPerceptron(Neuron): for j in rng.permutation(len(teach_data)): T = (j == T_NUMBER) ix = np.insert(teach_data[j][i].ravel(), 0, 1) - delta += ETA * (T - self.activation(ix.dot(self.weights))) * ix + err = T - self.activation(ix.dot(self.weights)) + delta += ETA * err * ix + self.errors.append(abs(err)) self.weights = self.weights + delta - # if np.linalg.norm(old_weights - self.weights) == 0.00: - # return self.weights - return self.weights + if np.linalg.norm(old_weights - self.weights) == 0.00: + return + return diff --git a/Reinforcement_Learning/Solution_Testing_1.py b/Reinforcement_Learning/Solution_Testing_1.py index 3e57838b5fa897a4f7726aaeb3e2339abe7daf12..eb5d8d30ef6ff4c5494a248d5470a99fa65b9e9d 100644 --- a/Reinforcement_Learning/Solution_Testing_1.py +++ b/Reinforcement_Learning/Solution_Testing_1.py @@ -1,3 +1,9 @@ +# README +# Threshold perceptron stagnates although not very accurate so the line +# for ending the training on the threshold early was commented out. +# Plot was added but commented out as the plots that happen too frequently cause an error. + + import numpy as np from prettytable import PrettyTable import Perceptrons @@ -10,7 +16,7 @@ trial_data = Perceptrons.test_data REPETION = 1 -def run_test(neuron, learning_rate): +def run_test(neuron, learning_rate, titel): global table table = PrettyTable() table.field_names = ["ETA", "0", "1", "2", "3", "4", "5", "6"] @@ -22,18 +28,19 @@ def run_test(neuron, learning_rate): res[i] = res[i] + round(abs(neuron.test(trial_array))) res = ["{:5d}".format(r) for r in res] table.add_row([ETA] + res) + # neuron.plot_errors(titel + f"{ETA}") print(table) return neuron -for i in range(7): +for i in range(6): Perceptrons.T_NUMBER = i E = np.array([0.05, 0.1, 0.2, 0.4, 0.75, 1, 2, 5]) print("Threshold Perceptron is looking for: ", Perceptrons.T_NUMBER) - run_test(Perceptrons.ThresholdPerceptron(20), E/4) + run_test(Perceptrons.ThresholdPerceptron(20), E/4, f"Threshold Perceptron digit {i} " ) print("Linear Perceptron is looking for: ", Perceptrons.T_NUMBER) - run_test(Perceptrons.LinearPerceptron(20), E / 160) + run_test(Perceptrons.LinearPerceptron(20), E / 160, f"Linear Perceptron digit {i} ") print("Sigmoid Gradient Descent Perceptron is looking for: ", Perceptrons.T_NUMBER) - x = run_test(Perceptrons.SGDPerceptron(20), E) + x = run_test(Perceptrons.SGDPerceptron(20), E, f"SGD Perceptron digit {i} ")