def build_feedforward_multilayered(input_number = 2, hidden_numbers = [2], output_number = 1, hidden_function = "tanh", output_function = "logsig"):
net = network()
net.num_of_inputs = input_number
inputs = range(input_number)
biasnode = net.bias_node
#Hidden layers
prev_layer = inputs
for hidden_number in hidden_numbers:
current_layer = []
for i in xrange(int(hidden_number)):
hidden = node(hidden_function)
connect_node(hidden, biasnode)
connect_nodes(hidden, prev_layer)
net.hidden_nodes.append(hidden)
current_layer.append(hidden)
prev_layer = current_layer
#Output nodes
for i in xrange(int(output_number)):
output = node(output_function)
connect_node(output, biasnode)
connect_nodes(output, prev_layer)
net.output_nodes.append(output)
return net
def crossover_node(mother, father):
child = network()
child.num_of_inputs = mother.num_of_inputs
tr = {}
for i in xrange(mother.num_of_inputs):
tr[i] = i
for bias in (mother.bias_node, father.bias_node):
tr[bias] = child.bias_node
for mother_node, father_node in zip(mother.hidden_nodes, father.hidden_nodes):
child_node = node(mother_node.activation_function, 1)
tr[mother_node] = child_node
tr[father_node] = child_node
#choose one node to pass on
_choice = sample([mother_node, father_node], 1)[0]
#map the weights and bias of that node to the child node
for keynode, weight in _choice.weights.items():
child_node.weights[tr[keynode]] = weight
child.hidden_nodes.append(child_node)
for mother_node, father_node in zip(mother.output_nodes, father.output_nodes):
child_node = node(mother_node.activation_function, 1)
tr[mother_node] = child_node
tr[father_node] = child_node
#choose one node to pass on
_choice = sample([mother_node, father_node], 1)[0]
#map the weights and bias of that node to the child node
for keynode, weight in _choice.weights.items():
child_node.weights[tr[keynode]] = weight
child.output_nodes.append(child_node)
return child
#glogger.debugPlot('Test Error\nMutation rate: ' + str(mutation_chance), best_error, style = 'r-', subset='best')
#glogger.debugPlot('Test Error\nMutation rate: ' + str(mutation_chance), error[top_networks[0]], style = 'b-', subset='top')
#Select the best 5 networks, mate them randomly and create 50 new networks
population = []
for child_index in xrange(population_size):
[mother, father] = sample(top_networks, 2)
if random() < cross_over_chance:
father = mother #Will create full mutation, no cross-over
population.append(network())
population[child_index].num_of_inputs = mother.num_of_inputs
bias_child = population[child_index].bias_node
#Hidden layer
for mother_node, father_node in zip(mother.hidden_nodes, father.hidden_nodes):
hidden = node(mother_node.activation_function, random_range)
weights = {}
for input_number in xrange(mother.num_of_inputs):
choice = sample([mother_node, father_node], 1)[0]
weights[input_number] = choice.weights[input_number]
choice = sample([(mother, mother_node), (father, father_node)], 1)[0]
weights[bias_child] = choice[1].weights[choice[0].bias_node]
_all = range(mother.num_of_inputs)
_all.append(bias_child)
connect_nodes(hidden, _all, weights)
population[child_index].hidden_nodes.append(hidden)
hidden_nodes = population[child_index].hidden_nodes
#Output nodes
