def setUp(self):
breeding_conf = dict(
{
"fitness_difference_threshold": 1,
"inherit_randomly_if_same_fitness_probability": 0.5,
"gene_inherited_as_disabled_probability": 0.5
}
)
self.breeder = Breeder(breeding_conf)
def dynamic_init(self):
"""
Initializes all parts of NEAT that are dependent upon
the client.
"""
self.init_db()
# Class containing huge configuration object.
# Loads config from JSON or uses default config.
self.config = NEATConfig(self._session["config_path"])
# selecting can mean:
# - selecting a single genome for mutation
# - selecting two genomes for breeding
# - selecting two clusters for combination
# - selecting two genomes from two given clusters for inter cluster
# breeding (since we don't really want to create ALL combinations)
self.selector = GenomeSelector(
self.genome_repository,
self.cluster_repository,
self.config.parameters["selection"]
)
# makes decisions lol
# things like what to do and stuff (breeding or mutation, if clustering
# is necessary etc)
self.decision_maker = DecisionMaker(
self.config.parameters["decision_making"]
)
# breeder creates a new genome from two given genomes
# it needs the gene_repository to register new genes and to look up used
# ones
self.breeder = Breeder(
self.config.parameters["breeding"]
)
# Mutator creates a new genome from a given genome
# it needs the gene_repository to register new genes and to look up used
# ones
self.mutator = Mutator(
self.gene_repository,
self.config.parameters["mutating"]
)
# Analyst analyzes a given genome and creates an AnalysisResult based on
# it
self.analyst = GenomeAnalyst()
# clusterer divides all existing and active genomes in clusters aka spe-
# cies
self.clusterer = GenomeClusterer(
self.genome_repository,
self.cluster_repository,
self.config.parameters["clustering"]
)
self.simulator = Simulator(self.gene_repository)
class TestBreeder(TestCase):
def setUp(self):
breeding_conf = dict(
{
"fitness_difference_threshold": 1,
"inherit_randomly_if_same_fitness_probability": 0.5,
"gene_inherited_as_disabled_probability": 0.5
}
)
self.breeder = Breeder(breeding_conf)
def test_breed_genomes(self):
fitter_genome = StorageGenome()
fitter_genome.fitness = 22.74
fitter_genome.genes = {
0: (True, 0.21),
1: (True, -0.56),
2: (True, 0.354),
3: (True, 0.98),
7: (True, 0.47),
8: (True, -0.13)
}
other_genome = StorageGenome()
other_genome.fitness = 19.233
other_genome.genes = {
0: (True, 0.1),
1: (True, 0.1),
2: (True, 0.1),
4: (True, 0.1),
5: (True, -0.1),
6: (True, -0.1)
}
similar_genome = StorageGenome()
similar_genome.fitness = 21.83
similar_genome.genes = {
0: (True, 0.1),
1: (True, 0.1),
2: (True, 0.1),
4: (True, 0.1),
5: (True, -0.1),
6: (True, -0.1)
}
new_genome = self.breeder.breed_genomes(
fitter_genome,
other_genome
)
new_genome_two = self.breeder.breed_genomes(
fitter_genome,
similar_genome
)
# print(new_genome.genes) # TODO: proper testing
# self.fail("Not implemented.")
self.assertListEqual(
list(new_genome.genes.keys()),
[0, 1, 2, 3, 7, 8],
"The wrong genes were inherited by breed_genomes()."
)
self.assertEqual(
new_genome.genes[3][1],
0.98,
"The differing gene with id 3 wasn't inherited correctly by breed_genomes()."
)
self.assertEqual(
new_genome.genes[7][1],
0.47,
"The differing gene with id 3 wasn't inherited correctly by breed_genomes()."
)
self.assertEqual(
new_genome.genes[8][1],
-0.13,
"The differing gene with id 3 wasn't inherited correctly by breed_genomes()."
)
in_set = [(g in new_genome_two.genes.keys()) for g in [0, 1, 2]]
self.assertTrue(
lambda : not False in in_set
)
differing_genes = [g for g in new_genome_two.genes.keys() if (g not in [0, 1, 2])]
for gene_id in differing_genes:
weight = new_genome_two.genes[gene_id][1]
parent_genome = fitter_genome \
if gene_id in fitter_genome.genes.keys() \
else similar_genome
original_weight = parent_genome.genes[gene_id][1]
self.assertEqual(
weight,
original_weight
)
class MainDirector(Director):
def __init__(self, **kwargs):
"""
:param kwargs:
- mode:
- exit: exits the program, default action if nothing is provided.
- new: creates a new database for a given simulation. requires parame-
ter simulation to be set.
- load: loads a database for a given simulation. requires parameter
simulation to be set.
- simulation: the name of the simulation that should be used. must cor-
respond to a module name in Simulation.
:return:
"""
self._maximum_timeouts = 5000
self.mode = kwargs.get('mode', 'exit')
self.selector = None # type: GenomeSelector
self.decision_maker = None # type: DecisionMaker
self.breeder = None # type: Breeder
self.mutator = None # type: Mutator
self.analyst = None # type: GenomeAnalyst
self.clusterer = None # type: GenomeClusterer
self.simulator = None # type: Simulator
self.simulation_connector = SimulationConnector() # type: SimulationConnector
self.database_connector = None # type: DatabaseConnector
self.gene_repository = None # type: GeneRepository
self.genome_repository = None # type: GenomeRepository
self.cluster_repository = None # type: ClusterRepository
self.config = None # type: NEATConfig
self._session = None # type: dict
self._discarded_genomes_count = 0 # type: int
if self.mode == 'exit':
exit()
elif self.mode == 'run_server':
startup_check = StartupCheck()
startup_check.run()
while True:
try:
self.idle() # TODO: exit command
except NetworkProtocolException as e:
print(e)
pass
def idle(self):
"""
Standard method that will be executed if local startup is done.
In this state, the Director will wait for the client.
"""
self._session = self.simulation_connector.get_session()
# Session tokens will identify a client.
# They can be useful for later parallelization.
# They also identify the database collections which will be used,
# so that different users can have their own storage and previous
# sessions can be loaded from storage.
self.dynamic_init() # This can be called after the client has connected
if debug:
self.database_connector.clear_collection("genomes")
self.database_connector.clear_collection("clusters")
self.database_connector.clear_collection("genes")
# In case of simulation run:
self.run()
def dynamic_init(self):
"""
Initializes all parts of NEAT that are dependent upon
the client.
"""
self.init_db()
# Class containing huge configuration object.
# Loads config from JSON or uses default config.
self.config = NEATConfig(self._session["config_path"])
# selecting can mean:
# - selecting a single genome for mutation
# - selecting two genomes for breeding
# - selecting two clusters for combination
# - selecting two genomes from two given clusters for inter cluster
# breeding (since we don't really want to create ALL combinations)
self.selector = GenomeSelector(
self.genome_repository,
self.cluster_repository,
self.config.parameters["selection"]
)
# makes decisions lol
# things like what to do and stuff (breeding or mutation, if clustering
# is necessary etc)
self.decision_maker = DecisionMaker(
self.config.parameters["decision_making"]
)
# breeder creates a new genome from two given genomes
# it needs the gene_repository to register new genes and to look up used
# ones
self.breeder = Breeder(
self.config.parameters["breeding"]
)
# Mutator creates a new genome from a given genome
# it needs the gene_repository to register new genes and to look up used
# ones
self.mutator = Mutator(
self.gene_repository,
self.config.parameters["mutating"]
)
# Analyst analyzes a given genome and creates an AnalysisResult based on
# it
self.analyst = GenomeAnalyst()
# clusterer divides all existing and active genomes in clusters aka spe-
# cies
self.clusterer = GenomeClusterer(
self.genome_repository,
self.cluster_repository,
self.config.parameters["clustering"]
)
self.simulator = Simulator(self.gene_repository)
def init_db(self):
# database connection is a connection to an arbitrary database that is
# used to store genes, genomes and nodes
self.database_connector = DatabaseConnector(
self._session["session_id"]
)
# gene_repository administrates all genes ever created
self.gene_repository = GeneRepository(
self.database_connector
)
# genome_repository administrates all genomes ever created
self.genome_repository = GenomeRepository(
self.database_connector
)
# cluster_repository administrates all clusters ever created
self.cluster_repository = ClusterRepository(
self.database_connector
)
def run(self):
"""
The main function where the simulation is run, new
genomes are created and discarded
This is where the evolutionary magic happens.
"""
# on new, creates random set of genomes based on configuration inside
# Simulation.given_simulation.config
self.decision_maker.reset_time()
# Init population if its not present yet.
if len(
list(self.genome_repository.get_current_population())
) < self.config.parameters["clustering"]["max_population"]:
self.init_population()
while True:
# 1. Simulation / wait for client
timeout_count = 0
advance_generation = None
while (timeout_count < self._maximum_timeouts) and \
advance_generation is None:
try:
advance_generation = self.perform_simulation_io()
except NetworkTimeoutException:
# TODO: log timeout event
timeout_count += 1
if not timeout_count < self._maximum_timeouts:
raise NetworkTimeoutException
# Either:
# * go on with loop, generate next generation
# * save database for later use, hand out session id to client
if not advance_generation:
print("Exiting...")
exit() # TODO: archive session / signal worker threads
# 2. Calculate offspring values
self.calculate_cluster_offspring()
# 3. Discarding / Regeneration
if self.decision_maker.inter_cluster_breeding_time:
# if it's time to cross-breed, first discard a few clusters
self.discard_clusters()
# then combine clusters
self.crossbreed_clusters()
else:
# if it's incest time, first discard a few genomes
self.discard_genomes()
# then refill the population
self.generate_new_genomes()
# 4. Advance time
self.decision_maker.advance_time()
def generate_new_genomes(self):
"""
Regenerates the population by selecting genomes for
mutation / breeding, running the generation process and performing analysis.
:return:
"""
mutation_percentage = self.decision_maker.mutation_percentage
genomes_for_mutation = self.selector.select_genomes_for_mutation(mutation_percentage)
genomes_for_breeding = self.selector.select_genomes_for_breeding(1 - mutation_percentage)
new_genomes = []
for genome in genomes_for_mutation:
new_genome = self.mutator.mutate_genome(genome)
new_genomes.append(new_genome)
for genome_one, genome_two in genomes_for_breeding:
new_genome = self.breeder.breed_genomes(
genome_one,
genome_two
)
new_genomes.append(new_genome)
for genome in new_genomes:
self.analyze_and_insert(genome)
def crossbreed_clusters(self):
"""
combines two clusters by breeding genomes of both clusters
:return:
"""
cluster_one, cluster_two = self.selector.select_clusters_for_combination()
for genome_one, genome_two in self.selector.select_cluster_combinations(
cluster_one,
cluster_two,
self._discarded_genomes_count
):
new_genome = self.breeder.breed_genomes(genome_one, genome_two)
self.analyze_and_insert(new_genome)
self._discarded_genomes_count = 0
def analyze_and_insert(self, genome: StorageGenome):
analysis_genome = AnalysisGenome(self.gene_repository, genome)
analysis_result = self.analyst.analyze(analysis_genome)
genome.analysis_result = analysis_result
self.genome_repository.insert_genome(genome)
self.clusterer.cluster_genome(genome)
def calculate_cluster_offspring(self):
"""
Calculates fitness values and offspring for clusters.
:return:
"""
self.clusterer.calculate_cluster_offspring_values()
def discard_genomes(self):
"""
discards a number of genomes
:return:
"""
for genome in self.selector.select_genomes_for_discarding():
self.genome_repository.disable_genome(genome.genome_id)
def discard_clusters(self):
"""
Discards a number of clusters
:return:
"""
for cluster in self.selector.select_clusters_for_discarding():
genomes_to_discard = self.genome_repository.get_genomes_in_cluster(cluster.cluster_id)
self._discarded_genomes_count += len(list(genomes_to_discard))
self.genome_repository.disable_genomes([i.genome_id for i in genomes_to_discard])
def perform_simulation_io(self):
genomes = list(self.genome_repository.get_current_population())
block_count = math.ceil(len(genomes) / self._session["block_size"])
genome_index = 0
fitness_values = {}
for block_id in range(block_count):
block = genomes[genome_index: genome_index + self._session["block_size"]]
self.simulation_connector.send_block(block, block_id)
block_inputs = self.simulation_connector.get_block_inputs(block_id)
self.simulation_connector.send_block_outputs(
self.compute_genome_outputs(block_inputs),
block_id
)
fitness_values = {
**fitness_values,
**self.simulation_connector.get_fitness_values(block_id)
}
genome_index += self._session["block_size"]
self.update_fitness_values(
fitness_values
)
return self.simulation_connector.get_advance_generation()
def compute_genome_outputs(
self,
block_inputs: Dict[ObjectId, Dict[str, float]]
) -> Dict[ObjectId, Dict[str, float]]:
results = dict({})
for genome_id, inputs in block_inputs.items():
storage_genome = self.genome_repository.get_genome_by_id(genome_id)
outputs = self.simulator.simulate_genome(storage_genome, inputs)
results[genome_id] = outputs
return results
def update_fitness_values(
self,
fitness_values: Dict[ObjectId, float]
) -> None:
for genome_id, fitness_value in fitness_values.items():
self.genome_repository.update_genome_fitness(
genome_id,
fitness_value
)
def init_population(self):
population_size = self.config.parameters["clustering"]["max_population"]
input_labels = self.config.parameters["genomes"]["inputs"]
output_labels = self.config.parameters["genomes"]["outputs"]
for i in range(population_size):
genome = StorageGenome(
inputs=input_labels,
outputs=output_labels
)
self.analyze_and_insert(genome)