aimacode · norvig · May 28, 2017 · Mar 29, 2017 · Mar 29, 2017
diff --git a/learning.py b/learning.py
@@ -3,7 +3,8 @@
 from utils import (
     removeall, unique, product, mode, argmax, argmax_random_tie, isclose, gaussian,
     dotproduct, vector_add, scalar_vector_product, weighted_sample_with_replacement,
-    weighted_sampler, num_or_str, normalize, clip, sigmoid, print_table, DataFile
+    weighted_sampler, num_or_str, normalize, clip, sigmoid, print_table,
+    DataFile, sigmoid_derivative
 )
 
 import copy
@@ -568,13 +569,17 @@ def predict(example):
     return predict
 
 
+def random_weights(min_value, max_value, num_weights):
+    return [random.uniform(min_value, max_value) for i in range(num_weights)]
+
+
 def BackPropagationLearner(dataset, net, learning_rate, epochs):
     """[Figure 18.23] The back-propagation algorithm for multilayer network"""
     # Initialise weights
     for layer in net:
         for node in layer:
-            node.weights = [random.uniform(-0.5, 0.5)
-                            for i in range(len(node.weights))]
+            node.weights = random_weights(min_value=-0.5, max_value=0.5,
+                                          num_weights=len(node.weights))
 
     examples = dataset.examples
     '''
@@ -612,10 +617,11 @@ def BackPropagationLearner(dataset, net, learning_rate, epochs):
             delta = [[] for i in range(n_layers)]
 
             # Compute outer layer delta
-            err = [t_val[i] - o_nodes[i].value
-                   for i in range(o_units)]
-            delta[-1] = [(o_nodes[i].value) * (1 - o_nodes[i].value) *
-                         (err[i]) for i in range(o_units)]
+
+            # Error for the MSE cost function
+            err = [t_val[i] - o_nodes[i].value for i in range(o_units)]
+            # The activation function used is the sigmoid function
+            delta[-1] = [sigmoid_derivative(o_nodes[i].value) * err[i] for i in range(o_units)]
 
             # Backward pass
             h_layers = n_layers - 2
@@ -624,11 +630,9 @@ def BackPropagationLearner(dataset, net, learning_rate, epochs):
                 h_units = len(layer)
                 nx_layer = net[i+1]
                 # weights from each ith layer node to each i + 1th layer node
-                w = [[node.weights[k] for node in nx_layer]
-                     for k in range(h_units)]
+                w = [[node.weights[k] for node in nx_layer] for k in range(h_units)]
 
-                delta[i] = [(layer[j].value) * (1 - layer[j].value) *
-                            dotproduct(w[j], delta[i+1])
+                delta[i] = [sigmoid_derivative(layer[j].value) * dotproduct(w[j], delta[i+1])
                             for j in range(h_units)]
 
             #  Update weights
@@ -754,7 +758,8 @@ def LinearLearner(dataset, learning_rate=0.01, epochs=100):
     X_col = [ones] + X_col
 
     # Initialize random weigts
-    w = [random.uniform(-0.5, 0.5) for _ in range(len(idx_i) + 1)]
+    num_weights = len(idx_i) + 1
+    w = random_weights(min_value=-0.5, max_value=0.5, num_weights=num_weights)
 
     for epoch in range(epochs):
         err = []
@@ -769,7 +774,6 @@ def LinearLearner(dataset, learning_rate=0.01, epochs=100):
         for i in range(len(w)):
             w[i] = w[i] + learning_rate * (dotproduct(err, X_col[i]) / num_examples)
 
-
     def predict(example):
         x = [1] + example
         return dotproduct(w, x)

diff --git a/tests/test_learning.py b/tests/test_learning.py
@@ -1,7 +1,7 @@
 from learning import parse_csv, weighted_mode, weighted_replicate, DataSet, \
                      PluralityLearner, NaiveBayesLearner, NearestNeighborLearner, \
                      NeuralNetLearner, PerceptronLearner, DecisionTreeLearner, \
-                     euclidean_distance, grade_learner, err_ratio
+                     euclidean_distance, grade_learner, err_ratio, random_weights
 from utils import DataFile
 
 
@@ -124,3 +124,18 @@ def test_perceptron():
 
     assert grade_learner(perceptron, tests) > 1/2
     assert err_ratio(perceptron, iris) < 0.4
+
+
+def test_random_weights():
+    min_value = -0.5
+    max_value = 0.5
+    num_weights = 10
+
+    test_weights = random_weights(min_value, max_value, num_weights)
+
+    assert len(test_weights) == num_weights
+
+    for weight in test_weights:
+        assert weight >= min_value and weight <= max_value
+
+
diff --git a/tests/test_utils.py b/tests/test_utils.py
@@ -154,6 +154,14 @@ def test_gaussian():
     assert gaussian(3,1,3) == 0.3989422804014327
 
 
+def test_sigmoid_derivative():
+    value = 1
+    assert sigmoid_derivative(value) == 0
+
+    value = 3
+    assert sigmoid_derivative(value) == -6
+
+
 def test_step():
     assert step(1) == step(0.5) == 1
     assert step(0) == 1

diff --git a/utils.py b/utils.py
@@ -249,6 +249,10 @@ def clip(x, lowest, highest):
     return max(lowest, min(x, highest))
 
 
+def sigmoid_derivative(value):
+    return value * (1 - value)
+
+
 def sigmoid(x):
     """Return activation value of x with sigmoid function"""
     return 1/(1 + math.exp(-x))