def main():
parser = argparse.ArgumentParser(description='Novelty Search.')
parser.add_argument('--config',
type=str,
help="yaml config file for ns",
default="")
args = parser.parse_args()
if not len(args.config):
raise Exception("You need to provide a yaml config file")
if len(args.config):
with open(args.config, "r") as fl:
config = yaml.load(fl, Loader=yaml.FullLoader)
if config["problem"]["name"] == "hardmaze":
max_steps = config["problem"]["max_steps"]
bd_type = config["problem"]["bd_type"]
assets = config["problem"]["assets"]
import HardMaze
problem = HardMaze.HardMaze(bd_type=bd_type,
max_steps=max_steps,
assets=assets)
elif config["problem"]["name"] == "large_ant_maze" or config[
"problem"]["name"] == "huge_ant_maze":
max_steps = config["problem"]["max_steps"]
bd_type = config["problem"]["bd_type"]
assets = config["problem"]["assets"]
pb_type = "huge" if config["problem"][
"name"] == "huge_ant_maze" else "large"
import LargeAntMaze
problem = LargeAntMaze.LargeAntMaze(pb_type=pb_type,
bd_type=bd_type,
max_steps=max_steps,
assets=assets)
else:
raise NotImplementedError("Problem type")
if config["novelty_estimator"]["type"] == "archive_based":
nov_estimator = NoveltyEstimators.ArchiveBasedNoveltyEstimator(
k=config["hyperparams"]["k"])
arch_types = {"list_based": Archives.ListArchive}
arch = arch_types[config["archive"]["type"]](
max_size=config["archive"]["max_size"],
growth_rate=config["archive"]["growth_rate"],
growth_strategy=config["archive"]["growth_strategy"],
removal_strategy=config["archive"]["removal_strategy"])
elif config["novelty_estimator"]["type"] == "learned":
bd_dims = problem.get_bd_dims()
embedding_dims = 2 * bd_dims
nov_estimator = NoveltyEstimators.LearnedNovelty1d(
in_dim=bd_dims,
emb_dim=embedding_dims,
pb_limits=problem.get_behavior_space_boundaries())
arch = None
if config["selector"]["type"] == "elitist_with_thresh":
selector = functools.partial(
MiscUtils.selBest, k=config["hyperparams"]["population_size"])
elif config["selector"]["type"] == "roulette_with_thresh":
roulette_msg = "Usage currently not supported: it ends up chosing the same element many times, this duplicates agent._ids etc"
roulette_msg += " fixing this bug is not a priority since selBest with thresholding actually works well"
raise Exception(roulette_msg)
elif config["selector"]["type"] == "nsga2_with_thresh":
selector = MiscUtils.NSGA2(
k=config["hyperparams"]["population_size"])
elif config["selector"]["type"] == "elitist":
selector = functools.partial(
MiscUtils.selBest,
k=config["hyperparams"]["population_size"],
automatic_threshold=False)
else:
raise NotImplementedError("selector")
in_dims = problem.dim_obs
out_dims = problem.dim_act
num_pop = config["hyperparams"]["population_size"]
if config["population"]["individual_type"] == "simple_fw_fc":
normalise_output_with = ""
num_hidden = 3
hidden_dim = 10
if "large_ant_maze" == config["problem"]["name"]:
normalise_output_with = "tanh"
num_hidden = 4
hidden_dim = 10
def make_ag():
return Agents.SmallFC_FW(
in_d=in_dims,
out_d=out_dims,
num_hidden=num_hidden,
hidden_dim=hidden_dim,
output_normalisation=normalise_output_with)
elif config["population"]["individual_type"] == "agent1d":
def make_ag():
return Agents.Agent1d(min(problem.env.phi_vals),
max(problem.env.phi_vals))
mutator_type = config["mutator"]["type"]
genotype_len = make_ag().get_genotype_len()
if mutator_type == "gaussian_same":
mutator_conf = config["mutator"]["gaussian_params"]
mu, sigma, indpb = mutator_conf["mu"], mutator_conf[
"sigma"], mutator_conf["indpb"]
mus = [mu] * genotype_len
sigmas = [sigma] * genotype_len
mutator = functools.partial(deap_tools.mutGaussian,
mu=mus,
sigma=sigmas,
indpb=indpb)
elif mutator_type == "poly_same":
mutator_conf = config["mutator"]["poly_params"]
eta, low, up, indpb = mutator_conf["eta"], mutator_conf[
"low"], mutator_conf["up"], mutator_conf["indpb"]
if config["population"]["individual_type"] == "agent1d":
dummy_ag = make_ag()
low = dummy_ag.min_val
up = dummy_ag.max_val
mutator = functools.partial(deap_tools.mutPolynomialBounded,
eta=eta,
low=low,
up=up,
indpb=indpb)
else:
raise NotImplementedError("mutation type")
map_t = "scoop" if config["use_scoop"] else "std"
visualise_bds = config["visualise_bds"]
ns = NoveltySearch(
archive=arch,
nov_estimator=nov_estimator,
mutator=mutator,
problem=problem,
selector=selector,
n_pop=num_pop,
n_offspring=config["hyperparams"]["offspring_size"],
agent_factory=make_ag,
visualise_bds_flag=visualise_bds,
map_type=map_t,
logs_root=config["ns_log_root"],
compute_parent_child_stats=config["compute_parent_child_stats"])
MiscUtils.bash_command(
["cp", args.config, ns.log_dir_path + "/config.yaml"])
stop_on_reaching_task = config["stop_when_task_solved"]
nov_estimator.log_dir = ns.log_dir_path
ns.disable_tqdm = config["disable_tqdm"]
ns.save_archive_to_file = config["archive"]["save_to_file"]
if ns.disable_tqdm:
print(
colored("[NS info] tqdm is disabled.",
"magenta",
attrs=["bold"]))
final_pop, solutions = ns(
iters=config["hyperparams"]["num_generations"],
stop_on_reaching_task=stop_on_reaching_task,
save_checkpoints=config["save_checkpoints"])
