def setUp(self):
neuron_sigma = 0.5
connection_sigma = 0.5
PS = NumericParamSpec
self.correct_types = set(
['Input', 'Simple', 'Sigmoid', 'Oscillator', 'default'])
self.correct_sigmoid_spec = GeneSpec("Sigmoid",
PS('bias', -1., 1., neuron_sigma),
PS('gain', 0., 1., neuron_sigma))
neuron_specs = [
GeneSpec("Input"),
self.correct_sigmoid_spec,
GeneSpec("Simple", PS('bias', -1., 1., neuron_sigma),
PS('gain', 0., 1., neuron_sigma)),
GeneSpec("Oscillator", PS("period", 0., 10., neuron_sigma),
PS("phase_offset", 0., 3.14, neuron_sigma),
PS("amplitude", 0., 10000., neuron_sigma)),
]
connection_specs = [
GeneSpec(
"default",
PS('weight', mutation_sigma=connection_sigma, mean_value=0.)),
]
self.netspec = NetworkSpec(neuron_specs, connection_specs)
def get_extended_mutation_spec(neuron_sigma, connection_sigma):
neuron_specs = [
GeneSpec("Input"),
GeneSpec("Sigmoid", PS('bias', -1., 1., neuron_sigma),
PS('gain', 0., 1., neuron_sigma)),
GeneSpec("Simple", PS('bias', -1., 1., neuron_sigma),
PS('gain', 0., 1., neuron_sigma)),
GeneSpec("Oscillator", PS("period", 0., 10., neuron_sigma),
PS("phase_offset", 0., 3.14, neuron_sigma),
PS("amplitude", 0., 10000., neuron_sigma)),
GeneSpec("Bias", PS('bias', -1., 1., neuron_sigma)),
GeneSpec("DifferentialCPG", PS('bias', -1., 1.)),
# these neurons (V-Neuron and X-Neuron) are for the nonlinear oscillator CPG model found in Ijspeert (2005):
GeneSpec("V-Neuron", PS('alpha', 0.05, 10., neuron_sigma),
PS('tau', 1., 50., neuron_sigma),
PS('energy', 0., 25., neuron_sigma)),
GeneSpec("X-Neuron", PS('tau', 0.01, 5.)),
]
connection_specs = [
GeneSpec("default",
PS('weight', mutation_sigma=connection_sigma, mean_value=0.)),
]
return NetworkSpec(neuron_specs, connection_specs)
def test_loopback_protection(self):
genome = self.loopback_test_genome.copy()
net_spec = NetworkSpec([], [GeneSpec('new_connection')])
mutator = Mutator(net_spec=net_spec)
con_added = False
con_added = mutator.add_connection_mutation(genome)
# print("---------------- ----------------")
# print(genome)
self.assertTrue(con_added, msg="Connection should have been added")
genome = self.loopback_test_genome.copy()
mutator = Mutator(net_spec=net_spec,
pure_input_types=('input_type_1, input_type_2'))
con_added = False
con_added = mutator.add_connection_mutation(genome)
# print("---------------- ----------------")
# print(genome)
self.assertFalse(con_added,
msg="Connection should have not been added")
def get_mutation_spec(neuron_sigma, connection_sigma):
neuron_specs = [
GeneSpec("Input"),
GeneSpec("Sigmoid", PS('bias', -1., 1., neuron_sigma),
PS('gain', 0., 1., neuron_sigma)),
GeneSpec("Simple", PS('bias', -1., 1., neuron_sigma),
PS('gain', 0., 1., neuron_sigma)),
GeneSpec("Oscillator", PS("period", 0., 10., neuron_sigma),
PS("phase_offset", 0., 3.14, neuron_sigma),
PS("amplitude", 0., 10000., neuron_sigma)),
]
connection_specs = [
GeneSpec("default",
PS('weight', mutation_sigma=connection_sigma, mean_value=0.)),
]
return NetworkSpec(neuron_specs, connection_specs)
def species_sizes(n, num_species):
return (n // num_species, ) * (num_species - 1) + (n // num_species +
n % num_species, )
def count_members(species_list):
return sum(len(species) for species in species_list)
## CREATE GENE SPECS ##
neuron_sigma = 0.25 # mutation sigma value for neuron params
conn_sigma = 10. # mutation sigma value for connection params
input_neuron_spec = GeneSpec(
'input',
PS('layer', gen.const('input')),
)
sigmoid_neuron_spec = GeneSpec(
'sigmoid',
PS('layer', gen.const('hidden')),
PS('bias', gen.uniform(), mut.gauss(neuron_sigma), bounds(-1., 1.)),
PS('gain', gen.uniform(), mut.gauss(neuron_sigma), bounds(0., 1.)),
)
connection_spec = GeneSpec(
'connection',
PS('weight', gen.gauss(0, conn_sigma), mut.gauss(conn_sigma)),
)
def produce_new_generation(pipeline, genome_fitness_list):
for _ in range(len(genome_fitness_list) - elite_num):
4, # size of the selection subsample (must be in the range [2, pop_size])
neuron_param_mut_proba=
0.5, # probability to mutate each single neuron in the genome
connection_param_mut_proba=
0.5, # probability to mutate each single connection in the genome
structural_augmentation_proba=
0.8, # probability to augment the topology of a newly created genome
structural_removal_proba=
0.0, # probability to diminish the topology of a newly created genome
speciation_threshold=
0.005 # genomes that are more similar than this value will be considered the same species
)
#### ###### #### #### #### #### #### #### #### #### #### #### #### #### #### #### #### #### #### #### ####
net_spec = NetworkSpec([
GeneSpec('input', NPS('layer', ['input'], mutable=False)),
GeneSpec('sigmoid', PS('bias', -1., 1., neuron_sigma, mutable=True),
PS('gain', 0, 1., neuron_sigma, mutable=True),
NPS('layer', ['hidden'], mutable=False))
], [
GeneSpec(
'default',
PS('weight', mutation_sigma=conn_sigma, mean_value=0., mutable=True))
])
mut = Mutator(net_spec, allowed_neuron_types=['sigmoid'])
neat_obj = NEAT(mutator=mut, **conf)
genome = neat_obj.get_init_genome(in1=neuron('input',
protected=True,