def run_evolution(sur_points, time_delta, space_delta, point_for_retrain, gens_to_change_fidelity, max_gens, pop_size,
archive_size, iter_id, deadline):
train_stations = [1, 2, 3]
initial_fidelity = (180, 56)
grid = CSVGridFile('../../samples/wind-exp-params-new.csv')
ww3_obs = \
[obs.time_series() for obs in
wave_watch_results(path_to_results='../../samples/ww-res/', stations=train_stations)]
train_model = FidelityFakeModel(grid_file=grid, observations=ww3_obs,
stations_to_out=train_stations, error=error_rmse_all,
forecasts_path='../../../2fidelity/*', forecasts_range=(0, 1),
sur_points=sur_points,
is_surrogate=True, initial_fidelity=initial_fidelity)
operators = default_operators()
handler = FidelityHandler(surrogates=train_model.surrogates_by_stations, time_delta=time_delta,
space_delta=space_delta,
point_for_retrain=point_for_retrain, gens_to_change_fidelity=gens_to_change_fidelity)
dyn_params = SPEA2.Params(max_gens=max_gens, pop_size=pop_size, archive_size=archive_size,
crossover_rate=0.7, mutation_rate=0.7,
mutation_value_rate=[0.1, 0.01, 0.001])
history, _, _ = DynamicSPEA2(
params=dyn_params,
objectives=partial(calculate_objectives_interp, train_model),
evolutionary_operators=operators,
fidelity_handler=handler).solution(verbose=True)
best = history.last()
fieldnames = ['ID', 'IterId', 'SetId', 'drf', 'cfw', 'stpm',
'rmse_all', 'rmse_peak', 'mae_all', 'mae_peak', 'deadline']
with open(os.path.join("C:\\metaopt2", "res-nondd.csv"), 'a', newline='') as csvfile:
writer = csv.DictWriter(csvfile, fieldnames=fieldnames)
row_to_write = {'ID': run_id, 'IterId': iter_id, 'SetId': 0,
'drf': best.genotype.drf,
'cfw': best.genotype.cfw,
'stpm': best.genotype.stpm,
'deadline': deadline}
metrics = all_error_metrics(best.genotype, models_to_tests)
for metric_name in metrics.keys():
stations_metrics = metrics[metric_name]
stations_to_write = {}
for station_idx in range(len(stations_metrics)):
key = f'st{station_idx + 1}'
stations_to_write.update({key: stations_metrics[station_idx]})
row_to_write.update({metric_name: stations_to_write})
writer.writerow(row_to_write)
return history.last().error_value
def run_evolution():
train_stations = [1]
grid = CSVGridFile('../../samples/wind-exp-params-new.csv')
ww3_obs = \
[obs.time_series() for obs in
wave_watch_results(path_to_results='../../samples/ww-res/', stations=train_stations)]
initial_fidelity = (180, 56)
sur_points = 5
train_model = FidelityFakeModel(grid_file=grid,
observations=ww3_obs,
stations_to_out=train_stations,
error=error_rmse_all,
forecasts_path='../../../2fidelity/*',
forecasts_range=(0, 1),
sur_points=sur_points,
is_surrogate=True,
initial_fidelity=initial_fidelity)
operators = default_operators()
handler = FidelityHandler(surrogates=train_model.surrogates_by_stations,
time_delta=15,
space_delta=7,
point_for_retrain=3,
gens_to_change_fidelity=10)
dyn_params = SPEA2.Params(max_gens=100,
pop_size=10,
archive_size=5,
crossover_rate=0.7,
mutation_rate=0.7,
mutation_value_rate=[0.1, 0.01, 0.001])
ex_time = DynamicSPEA2PerfModel.get_execution_time(sur_points,
initial_fidelity,
dyn_params, handler)
print(ex_time)
_, _, points_by_fid = DynamicSPEA2(
params=dyn_params,
objectives=partial(calculate_objectives_interp, train_model),
evolutionary_operators=operators,
fidelity_handler=handler,
points_by_fidelity=default_points_by_fidelity(size=5)).solution(
verbose=True)
points_by_fid
def run_genetic_opt(max_gens, pop_size, archive_size, crossover_rate,
mutation_rate, mutation_value_rate, sur_points, stations,
**kwargs):
grid = CSVGridFile('../../samples/wind-exp-params-new.csv')
ww3_obs = \
[obs.time_series() for obs in wave_watch_results(path_to_results='../../samples/ww-res/', stations=stations)]
train_model = FidelityFakeModel(grid_file=grid,
observations=ww3_obs,
stations_to_out=stations,
error=error_rmse_all,
forecasts_path='../../../2fidelity/*',
is_surrogate=True,
sur_points=sur_points,
initial_fidelity=(180, 56))
test_range = (0, 1)
test_model = model_all_stations(forecasts_range=test_range)
operators = default_operators()
handler = FidelityHandler(surrogates=train_model.surrogates_by_stations,
time_delta=30,
space_delta=14,
point_for_retrain=3,
gens_to_change_fidelity=10)
history, archive_history, _ = DynamicSPEA2(
params=SPEA2.Params(max_gens=max_gens,
pop_size=30,
archive_size=10,
crossover_rate=crossover_rate,
mutation_rate=mutation_rate,
mutation_value_rate=mutation_value_rate),
objectives=partial(calculate_objectives_interp, train_model),
evolutionary_operators=operators,
fidelity_handler=handler).solution(verbose=False)
exptime2 = str(datetime.datetime.now().time()).replace(":", "-")
# save_archive_history(archive_history, f'rob-exp-bl-{exptime2}.csv')
params = history.last().genotype
return history.last()
def objective(args):
sur_points, gens_to_change_fidelity, max_gens, pop_size = args
sur_points = int(round(sur_points))
time_delta = 30
space_delta = 14
pop_size = int(round(pop_size))
archive_size = round(pop_size / 2)
point_for_retrain = round(archive_size / 2)
# int(round(point_for_retrain))
gens_to_change_fidelity = int(round(gens_to_change_fidelity))
handler = FidelityHandler(surrogates=[], time_delta=15, space_delta=7,
point_for_retrain=3, gens_to_change_fidelity=10)
initial_fidelity = (180, 56)
dyn_params = SPEA2.Params(max_gens=100, pop_size=10, archive_size=5,
crossover_rate=0.7, mutation_rate=0.7,
mutation_value_rate=[0.1, 0.01, 0.001])
ex_time = DynamicSPEA2PerfModel.get_execution_time(sur_points, initial_fidelity, dyn_params, handler)
if (ex_time >= deadline or ex_time < deadline * 0.7):
print(ex_time)
return 99999
print("OBJ", sur_points, point_for_retrain, gens_to_change_fidelity)
score = run_evolution(sur_points, time_delta, space_delta, point_for_retrain, gens_to_change_fidelity, max_gens,
pop_size, archive_size)
if (len(best_score_history) == 0 or score < best_score_history[len(best_score_history) - 1][
0]):
best_score = score
best_score_history.append([best_score, ])
score_history.append([score, ])
print("SCORE FOUND")
print(score)
return score
def run_evolution(sur_points, time_delta, space_delta, point_for_retrain,
gens_to_change_fidelity, max_gens, pop_size, archive_size):
train_stations = [1]
initial_fidelity = (180, 56)
#sur_points = 5
dyn_params = SPEA2.Params(max_gens=max_gens,
pop_size=pop_size,
archive_size=archive_size,
crossover_rate=0.7,
mutation_rate=0.7,
mutation_value_rate=[0.1, 0.01, 0.001])
grid = CSVGridFile('../../samples/wind-exp-params-new.csv')
ww3_obs = \
[obs.time_series() for obs in
wave_watch_results(path_to_results='../../samples/ww-res/', stations=train_stations)]
train_model = FidelityFakeModel(grid_file=grid,
observations=ww3_obs,
stations_to_out=train_stations,
error=error_rmse_all,
forecasts_path='../../../2fidelity/*',
forecasts_range=(0, 1),
sur_points=sur_points,
is_surrogate=True,
initial_fidelity=initial_fidelity)
operators = default_operators()
handler = FidelityHandler(surrogates=train_model.surrogates_by_stations,
time_delta=15,
space_delta=7,
point_for_retrain=3,
gens_to_change_fidelity=10)
dyn_params = SPEA2.Params(max_gens=max_gens,
pop_size=pop_size,
archive_size=archive_size,
crossover_rate=0.7,
mutation_rate=0.7,
mutation_value_rate=[0.1, 0.01, 0.001])
# ex_time = DynamicSPEA2PerfModel.get_execution_time(sur_points, initial_fidelity, dyn_params, handler)
# print(ex_time)
history, _, _ = DynamicSPEA2(
params=dyn_params,
objectives=partial(calculate_objectives_interp, train_model),
evolutionary_operators=operators,
fidelity_handler=handler).solution(verbose=True)
fieldnames = [
'ID', 'deadline', 'sur_points', 'gens_to_change_fidelity', 'max_gens',
'pop_size', 'error'
]
with open(os.path.join("C:\\metaopt1", "res1.csv"), 'a',
newline='') as csvfile:
writer = csv.DictWriter(csvfile, fieldnames=fieldnames)
row_to_write = {
'ID': 0,
'deadline': deadline,
'sur_points': sur_points,
'gens_to_change_fidelity': gens_to_change_fidelity,
'max_gens': max_gens,
'pop_size': pop_size,
'error': history.last().error_value
}
writer.writerow(row_to_write)
return history.last().error_value
def objective(args):
sur_points, gens_to_change_fidelity, max_gens, pop_size = args
sur_points = int(round(sur_points))
time_delta = 30
space_delta = 14
pop_size = int(round(pop_size))
archive_size = round(pop_size / 2)
point_for_retrain = round(archive_size / 2)
# int(round(point_for_retrain))
gens_to_change_fidelity = int(round(gens_to_change_fidelity))
handler = FidelityHandler(surrogates=[],
time_delta=15,
space_delta=7,
point_for_retrain=3,
gens_to_change_fidelity=10)
initial_fidelity = (180, 56)
dyn_params = SPEA2.Params(max_gens=max_gens,
pop_size=pop_size,
archive_size=archive_size,
crossover_rate=0.7,
mutation_rate=0.7,
mutation_value_rate=[0.1, 0.01, 0.001])
ex_time = DynamicSPEA2PerfModel.get_execution_time(sur_points,
initial_fidelity,
dyn_params, handler)
if (ex_time >= deadline or ex_time < deadline * 0.7):
print(ex_time)
return 99999
print("OBJ", sur_points, point_for_retrain, gens_to_change_fidelity)
score = run_evolution(sur_points, time_delta, space_delta,
point_for_retrain, gens_to_change_fidelity, max_gens,
pop_size, archive_size)
if (len(best_score_history) == 0
or score < best_score_history[len(best_score_history) - 1][0]):
best_score = score
best_score_history.append([
best_score,
])
score_history.append([
score,
])
print("SCORE FOUND")
print(score)
return score
# optimize
fieldnames = [
'ID', 'deadline', 'sur_points', 'gens_to_change_fidelity', 'max_gens',
'pop_size', 'error'
]
# run_id+=1
writer = csv.DictWriter(os.path.join("C:\\metaopt1", "res1.csv"),
'a',
newline='',
fieldnames=fieldnames)
writer.writeheader()
pop_size = 10 + 5 * (deadline_ratio - 1)
max_gens = 10 + 5 * (deadline_ratio - 1)
time_delta = 30
space_delta = 14
pop_size = int(round(pop_size))
archive_size = round(pop_size / 2)
point_for_retrain = round(archive_size / 4)
gens_to_change_fidelity = round(max_gens / 2)
handler = FidelityHandler(surrogates=[], time_delta=time_delta, space_delta=space_delta,
point_for_retrain=point_for_retrain,
gens_to_change_fidelity=gens_to_change_fidelity)
dyn_params = SPEA2.Params(max_gens=max_gens, pop_size=pop_size, archive_size=archive_size,
crossover_rate=0.7, mutation_rate=0.7,
mutation_value_rate=[0.1, 0.01, 0.001])
ex_time = DynamicSPEA2PerfModel.get_execution_time(initial_surrogate_points, initial_fidelity, dyn_params,
handler)
wasted_time = ex_time
initial_surrogate_points += 50
if wasted_time >= deadline:
evo_res = run_evolution(initial_surrogate_points - 50, time_delta, space_delta, point_for_retrain,
gens_to_change_fidelity, max_gens - 5, pop_size - 5,
archive_size, iterId, deadline)