def main():
# make the scheduler first
with Scheduler({'nanny': True, 'interface': 'lo'}) as mpi_scheduler:
# create and run GA
# acs = AdaptiveCoolingSchedule(1, 0.2, 0, 10000, 1.0, 20)
# ga = ACSGa(initial_state=np.array([0.4, 0.3, -0.25, 0.01]),
# population_size=5,
# scheduler=mpi_scheduler,
# global_seed=96879,
# cooling_schedule=acs,
# table_size=20000,
# max_table_step=1,
# member_type=Member,
# save_filename=Path("test.acsga"),
# save_every=1000)
# ga.run(sphere, 350, take_member=False)
acs = AdaptiveCoolingSchedule(5.0, 100.0, 0.0, 500, 1.0, 50) # 20
ga = ACSGa(initial_state=np.array([0.4, 0.3, -0.25, 0.01]),
population_size=5,
scheduler=mpi_scheduler,
global_seed=96879,
cooling_schedule=acs,
table_size=20000,
max_table_step=1,
member_type=Member,
save_filename=Path("test.acsga"),
save_every=1000)
ga.run(rosenbrock, 2000, take_member=False)
def main():
# make the scheduler first
with Scheduler({'nanny': True, 'interface': 'lo'}) as mpi_scheduler:
# create and run GA
ga = AsyncCSAGa(initial_state=np.array([0.4, 0.3, -0.25, 0.01]),
initial_sigma=np.ones(4) * 5,
population_size=5,
scheduler=mpi_scheduler,
global_seed=96879,
table_size=200000,
max_table_step=1,
member_type=DiagnosticCSAMember,
save_filename=Path("test.csaga"))
ga.run(sphere, 1000, take_member=False)
def main():
# make the scheduler first
with Scheduler({'nanny': True, 'interface': 'lo'}) as mpi_scheduler:
# create and run GA
ga = AsyncCSAGa(initial_state=np.array([0.4, 0.3, -0.25, 0.01]),
initial_sigma=np.ones(4) * 5,
population_size=5,
scheduler=mpi_scheduler,
global_seed=96879,
table_size=200000,
max_table_step=1,
member_type=CSAMember,
save_filename=Path("test.csaga"),
save_every=1000)
ga.run(sphere, 300, take_member=False)
# load pop from file and continue
ga = AsyncCSAGa.from_file("test.csaga",
scheduler=mpi_scheduler,
global_seed=432,
save_filename="test.csaga")
ga.run(sphere, 50, take_member=False)
def run(self,
fitness_function: Callable[[np.ndarray], float],
num_iterations: int,
fitness_kwargs: Dict = None,
take_member: bool = False):
"""
Executes the genetic algorithm.
:param fitness_function: a function that returns the fitness of a lineage.
:param num_iterations: the number of steps to run.
:param fitness_kwargs: any key word arguments for fitness_function.
:param take_member: whether the fitness function requires the member to be
provided or not (if not then expects an array) (default: False).
"""
if fitness_kwargs is None:
fitness_kwargs = {}
if self._save_every is None:
self._save_every = num_iterations
# distribute members for each worker
self._scheduler.wait_on_workers()
num_workers = self._scheduler.num_workers()
if num_workers == 0:
raise ValueError("Error: there are no workers.")
workers = self._scheduler.get_worker_names()
if len(workers) == 0:
raise ValueError("Error: there are no workers.")
members = [
self._scheduler.client.submit(initialize_member,
self._member_type,
self._member_type_kwargs,
workers=[worker])
for worker in workers
]
# skip if population is loaded from file
if not self._from_file:
# create initial batch
initial_batch = [
self._scheduler.client.submit(dispatch_work,
fitness_function,
self._population[pop]['lineage'],
members[i],
i,
fitness_kwargs,
take_member,
workers=[workers[i]])
for i, pop_group in enumerate(
split_work(list(range(self._population_size)),
num_workers)) for pop in pop_group
]
# wait for initial batch to complete and fill initial population
for completed_job in Scheduler.as_completed(initial_batch):
fitness, lineage, index = completed_job.result()
self._set_initial(lineage, fitness)
# submit jobs to all workers or till num iterations is saturated
jobs = []
for index in range(min(num_iterations, len(members))):
self._anneal()
jobs.append(
self._scheduler.client.submit(dispatch_work,
fitness_function,
self._mutation(
self._selection()),
members[index],
index,
fitness_kwargs,
take_member=take_member,
workers=[workers[index]]))
self._step += 1
# iterate ga until num_iterations reached
working_batch = Scheduler.as_completed(jobs)
for completed_job in working_batch:
fitness, lineage, index = completed_job.result()
self._replacement(lineage, fitness)
self._update_fitness_history()
if (self._save_filename is not None) and \
(self._step % self._save_every == 0):
self.save_population(self._save_filename)
if self._step < num_iterations:
self._anneal()
working_batch.add(
self._scheduler.client.submit(dispatch_work,
fitness_function,
self._mutation(
self._selection()),
members[index],
index,
fitness_kwargs,
take_member=take_member,
workers=[workers[index]]))
self._step += 1
if self._save_filename is not None:
self.save_population(self._save_filename)