def train_and_eval(trial: optuna.Trial, study_dir: str, seed: int):
"""
Objective function for the Optuna `Study` to maximize.
.. note::
Optuna expects only the `trial` argument, thus we use `functools.partial` to sneak in custom arguments.
:param trial: Optuna Trial object for hyper-parameter optimization
:param study_dir: the parent directory for all trials in this study
:param seed: seed value for the random number generators, pass `None` for no seeding
:return: objective function value
"""
# Synchronize seeds between Optuna trials
pyrado.set_seed(seed)
# Environment
env_hparams = dict(dt=1 / 250., max_steps=1500)
env = QQubeSwingUpSim(**env_hparams)
env = ActNormWrapper(env)
# Policy
policy_hparam = dict(feats=FeatureStack([
identity_feat, sign_feat, abs_feat, squared_feat, cubic_feat,
ATan2Feat(1, 2),
MultFeat([4, 5])
]))
policy = LinearPolicy(spec=env.spec, **policy_hparam)
# Algorithm
algo_hparam = dict(
num_workers=1, # parallelize via optuna n_jobs
max_iter=50,
pop_size=trial.suggest_int('pop_size', 50, 200),
num_rollouts=trial.suggest_int('num_rollouts', 4, 10),
num_is_samples=trial.suggest_int('num_is_samples', 5, 40),
expl_std_init=trial.suggest_uniform('expl_std_init', 0.1, 0.5),
symm_sampling=trial.suggest_categorical('symm_sampling',
[True, False]),
)
csv_logger = create_csv_step_logger(
osp.join(study_dir, f'trial_{trial.number}'))
algo = PoWER(osp.join(study_dir, f'trial_{trial.number}'),
env,
policy,
**algo_hparam,
logger=csv_logger)
# Train without saving the results
algo.train(snapshot_mode='latest', seed=seed)
# Evaluate
min_rollouts = 1000
sampler = ParallelRolloutSampler(
env, policy, num_workers=1,
min_rollouts=min_rollouts) # parallelize via optuna n_jobs
ros = sampler.sample()
mean_ret = sum([r.undiscounted_return() for r in ros]) / min_rollouts
return mean_ret
def test_bayrn_power(ex_dir, env: SimEnv, bayrn_hparam: dict):
pyrado.set_seed(0)
# Environments and domain randomization
env_real = deepcopy(env)
env_sim = DomainRandWrapperLive(env, create_zero_var_randomizer(env))
dp_map = create_default_domain_param_map_qq()
env_sim = MetaDomainRandWrapper(env_sim, dp_map)
env_real.domain_param = dict(mass_pend_pole=0.024 * 1.1,
mass_rot_pole=0.095 * 1.1)
env_real = wrap_like_other_env(env_real, env_sim)
# Policy and subroutine
policy_hparam = dict(energy_gain=0.587, ref_energy=0.827)
policy = QQubeSwingUpAndBalanceCtrl(env_sim.spec, **policy_hparam)
subrtn_hparam = dict(
max_iter=1,
pop_size=8,
num_init_states_per_domain=1,
num_is_samples=4,
expl_std_init=0.1,
num_workers=1,
)
subrtn = PoWER(ex_dir, env_sim, policy, **subrtn_hparam)
# Set the boundaries for the GP
dp_nom = inner_env(env_sim).get_nominal_domain_param()
ddp_space = BoxSpace(
bound_lo=np.array([
0.8 * dp_nom["mass_pend_pole"], 1e-8,
0.8 * dp_nom["mass_rot_pole"], 1e-8
]),
bound_up=np.array([
1.2 * dp_nom["mass_pend_pole"], 1e-7,
1.2 * dp_nom["mass_rot_pole"], 1e-7
]),
)
# Create algorithm and train
algo = BayRn(ex_dir,
env_sim,
env_real,
subrtn,
ddp_space,
**bayrn_hparam,
num_workers=1)
algo.train()
assert algo.curr_iter == algo.max_iter or algo.stopping_criterion_met()
env = DomainRandWrapperLive(env, randomizer)
# Policy
bounds = ([0.0, 0.25, 0.5], [1.0, 1.5, 2.5])
policy_hparam = dict(rbf_hparam=dict(num_feat_per_dim=9, bounds=bounds, scale=None), dim_mask=2)
policy = DualRBFLinearPolicy(env.spec, **policy_hparam)
# Algorithm
algo_hparam = dict(
max_iter=15,
pop_size=100,
num_is_samples=10,
num_init_states_per_domain=2,
num_domains=10,
expl_std_init=np.pi / 12,
expl_std_min=0.02,
num_workers=8,
)
algo = PoWER(ex_dir, env, policy, **algo_hparam)
# Save the hyper-parameters
save_dicts_to_yaml(
dict(env=env_hparams, seed=args.seed),
dict(policy=policy_hparam),
dict(algo=algo_hparam, algo_name=algo.name),
save_dir=ex_dir,
)
# Jeeeha
algo.train(seed=args.seed, snapshot_mode="best")
env_real = wrap_like_other_env(env_real, env_sim)
# PoWER + energy-based controller setup
policy_hparam = dict(energy_gain=0.587, ref_energy=0.827, acc_max=10.)
policy = QQubeSwingUpAndBalanceCtrl(env_sim.spec, **policy_hparam)
subrtn_hparam = dict(
max_iter=10,
pop_size=50,
num_rollouts=8,
num_is_samples=5,
expl_std_init=2.0,
expl_std_min=0.02,
symm_sampling=False,
num_workers=4,
)
subrtn = PoWER(ex_dir, env_sim, policy, **subrtn_hparam)
# PoWER + linear policy setup
# policy_hparam = dict(
# feats=FeatureStack([identity_feat, sign_feat, abs_feat, squared_feat,
# MultFeat([2, 5]), MultFeat([3, 5]), MultFeat([4, 5])])
# )
# policy = LinearPolicy(spec=env_sim.spec, **policy_hparam)
# subrtn_hparam = dict(
# max_iter=20,
# pop_size=200,
# num_rollouts=6,
# num_is_samples=10,
# expl_std_init=2.0,
# expl_std_min=0.02,
# symm_sampling=False,