def train_argmax_policy(
load_dir: pyrado.PathLike,
env_sim: MetaDomainRandWrapper,
subrtn: Algorithm,
num_restarts: int,
num_samples: int,
policy_param_init: to.Tensor = None,
valuefcn_param_init: to.Tensor = None,
subrtn_snapshot_mode: str = "best",
) -> Policy:
"""
Train a policy based on the maximizer of the posterior mean.
:param load_dir: directory to load from
:param env_sim: simulation environment
:param subrtn: algorithm which performs the policy / value-function optimization
:param num_restarts: number of restarts for the optimization of the acquisition function
:param num_samples: number of samples for the optimization of the acquisition function
:param policy_param_init: initial policy parameter values for the subroutine, set `None` to be random
:param valuefcn_param_init: initial value function parameter values for the subroutine, set `None` to be random
:param subrtn_snapshot_mode: snapshot mode for saving during training of the subroutine
:return: the final BayRn policy
"""
# Load the required data
cands = pyrado.load("candidates.pt", load_dir)
cands_values = pyrado.load("candidates_values.pt",
load_dir).unsqueeze(1)
ddp_space = pyrado.load("ddp_space.pkl", load_dir)
if cands.shape[0] > cands_values.shape[0]:
print_cbt(
f"There are {cands.shape[0]} candidates but only {cands_values.shape[0]} evaluations. Ignoring the"
f"candidates without evaluation for computing the argmax.",
"y",
)
cands = cands[:cands_values.shape[0], :]
# Find the maximizer
argmax_cand = BayRn.argmax_posterior_mean(cands, cands_values,
ddp_space, num_restarts,
num_samples)
# Set the domain randomizer
env_sim.adapt_randomizer(argmax_cand.numpy())
# Reset the subroutine algorithm which includes resetting the exploration
subrtn.reset()
# Do a warm start
subrtn.init_modules(warmstart=True,
policy_param_init=policy_param_init,
valuefcn_param_init=valuefcn_param_init)
subrtn.train(snapshot_mode=subrtn_snapshot_mode,
meta_info=dict(suffix="argmax"))
return subrtn.policy
def train_argmax_policy(load_dir: str,
env_sim: MetaDomainRandWrapper,
subroutine: Algorithm,
num_restarts: int,
num_samples: int,
policy_param_init: to.Tensor = None,
valuefcn_param_init: to.Tensor = None) -> Policy:
"""
Train a policy based on the maximizer of the posterior mean.
:param load_dir: directory to load from
:param env_sim: simulation environment
:param subroutine: algorithm which performs the policy / value-function optimization
:param num_restarts: number of restarts for the optimization of the acquisition function
:param num_samples: number of samples for the optimization of the acquisition function
:param policy_param_init: initial policy parameter values for the subroutine, set `None` to be random
:param valuefcn_param_init: initial value function parameter values for the subroutine, set `None` to be random
:return: the final BayRn policy
"""
# Load the required data
cands = to.load(osp.join(load_dir, 'candidates.pt'))
cands_values = to.load(osp.join(load_dir,
'candidates_values.pt')).unsqueeze(1)
bounds = to.load(osp.join(load_dir, 'bounds.pt'))
uc_normalizer = UnitCubeProjector(bounds[0, :], bounds[1, :])
# Find the maximizer
argmax_cand = BayRn.argmax_posterior_mean(cands, cands_values,
uc_normalizer, num_restarts,
num_samples)
# Set the domain randomizer given the hyper-parameters
env_sim.adapt_randomizer(argmax_cand.numpy())
# Reset the subroutine's algorithm which includes resetting the exploration
subroutine.reset()
# Reset the subroutine's policy (and value function)
subroutine.policy.init_param(policy_param_init)
if isinstance(subroutine, ActorCritic):
subroutine.critic.value_fcn.init_param(valuefcn_param_init)
if policy_param_init is None:
print_cbt('Learning the argmax solution from scratch', 'y')
else:
print_cbt('Learning the argmax solution given an initialization',
'y')
subroutine.train(
snapshot_mode='best') # meta_info=dict(prefix='final')
return subroutine.policy
def create_qqsu_setup():
# Environments
env_hparams = dict(dt=1 / 100.0, max_steps=600)
env_real = QQubeSwingUpSim(**env_hparams)
env_real.domain_param = dict(
mass_rot_pole=0.095 * 0.9, # 0.095*0.9 = 0.0855
mass_pend_pole=0.024 * 1.1, # 0.024*1.1 = 0.0264
length_rot_pole=0.085 * 0.9, # 0.085*0.9 = 0.0765
length_pend_pole=0.129 * 1.1, # 0.129*1.1 = 0.1419
)
env_sim = QQubeSwingUpSim(**env_hparams)
randomizer = DomainRandomizer(
NormalDomainParam(name="mass_rot_pole",
mean=0.0,
std=1e-9,
clip_lo=1e-3),
NormalDomainParam(name="mass_pend_pole",
mean=0.0,
std=1e-9,
clip_lo=1e-3),
NormalDomainParam(name="length_rot_pole",
mean=0.0,
std=1e-9,
clip_lo=1e-3),
NormalDomainParam(name="length_pend_pole",
mean=0.0,
std=1e-9,
clip_lo=1e-3),
)
env_sim = DomainRandWrapperLive(env_sim, randomizer)
dp_map = {
0: ("mass_rot_pole", "mean"),
1: ("mass_rot_pole", "std"),
2: ("mass_pend_pole", "mean"),
3: ("mass_pend_pole", "std"),
4: ("length_rot_pole", "mean"),
5: ("length_rot_pole", "std"),
6: ("length_pend_pole", "mean"),
7: ("length_pend_pole", "std"),
}
# trafo_mask = [False, True, False, True, False, True, False, True]
trafo_mask = [True] * 8
env_sim = MetaDomainRandWrapper(env_sim, dp_map)
# Policies (the behavioral policy needs to be deterministic)
behavior_policy = QQubeSwingUpAndBalanceCtrl(env_sim.spec)
prior = DomainRandomizer(
NormalDomainParam(name="mass_rot_pole", mean=0.095, std=0.095 / 10),
NormalDomainParam(name="mass_pend_pole", mean=0.024, std=0.024 / 10),
NormalDomainParam(name="length_rot_pole", mean=0.085, std=0.085 / 10),
NormalDomainParam(name="length_pend_pole", mean=0.129, std=0.129 / 10),
)
ddp_policy = DomainDistrParamPolicy(mapping=dp_map,
trafo_mask=trafo_mask,
prior=prior,
scale_params=False)
return env_sim, env_real, env_hparams, dp_map, behavior_policy, ddp_policy
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()
def create_bob_setup():
# Environments
env_hparams = dict(dt=1 / 100., max_steps=500)
env_real = BallOnBeamSim(**env_hparams)
env_real.domain_param = dict(
# l_beam=1.95,
# ang_offset=-0.03,
g=10.81)
env_sim = BallOnBeamSim(**env_hparams)
randomizer = DomainRandomizer(
# NormalDomainParam(name='l_beam', mean=0, std=1e-12, clip_lo=1.5, clip_up=3.5),
# UniformDomainParam(name='ang_offset', mean=0, halfspan=1e-12),
NormalDomainParam(name='g', mean=0, std=1e-12), )
env_sim = DomainRandWrapperLive(env_sim, randomizer)
dp_map = {
# 0: ('l_beam', 'mean'), 1: ('l_beam', 'std'),
# 2: ('ang_offset', 'mean'), 3: ('ang_offset', 'halfspan')
0: ('g', 'mean'),
1: ('g', 'std')
}
env_sim = MetaDomainRandWrapper(env_sim, dp_map)
# Policies (the behavioral policy needs to be deterministic)
behavior_policy = LinearPolicy(env_sim.spec,
feats=FeatureStack(
[identity_feat, sin_feat]))
behavior_policy.param_values = to.tensor(
[3.8090, -3.8036, -1.0786, -2.4510, -0.9875, -1.3252, 3.1503, 1.4443])
prior = DomainRandomizer(
# NormalDomainParam(name='l_beam', mean=2.05, std=2.05/10),
# UniformDomainParam(name='ang_offset', mean=0.03, halfspan=0.03/10),
NormalDomainParam(name='g', mean=8.81, std=8.81 / 10), )
# trafo_mask = [False, True, False, True]
trafo_mask = [True, True]
ddp_policy = DomainDistrParamPolicy(mapping=dp_map,
trafo_mask=trafo_mask,
prior=prior,
scale_params=True)
return env_sim, env_real, env_hparams, dp_map, behavior_policy, ddp_policy
std=1e6,
clip_lo=1e-3),
)
env_sim = DomainRandWrapperLive(env_sim, randomizer)
dp_map = {
0: ("mass_rot_pole", "mean"),
1: ("mass_rot_pole", "std"),
2: ("mass_pend_pole", "mean"),
3: ("mass_pend_pole", "std"),
4: ("length_rot_pole", "mean"),
5: ("length_rot_pole", "std"),
6: ("length_pend_pole", "mean"),
7: ("length_pend_pole", "std"),
}
trafo_mask = [False, True, False, True, False, True, False, True]
env_sim = MetaDomainRandWrapper(env_sim, dp_map)
# Subroutine for policy improvement
behav_policy_hparam = dict(hidden_sizes=[64, 64], hidden_nonlin=to.tanh)
behav_policy = FNNPolicy(spec=env_sim.spec, **behav_policy_hparam)
vfcn_hparam = dict(hidden_sizes=[32, 32], hidden_nonlin=to.relu)
vfcn = FNNPolicy(spec=EnvSpec(env_sim.obs_space, ValueFunctionSpace),
**vfcn_hparam)
critic_hparam = dict(
gamma=0.9844224855479998,
lamda=0.9700148505302241,
num_epoch=5,
batch_size=500,
standardize_adv=False,
lr=7.058326426522811e-4,
max_grad_norm=6.0,
def test_simopt_cem_ppo(ex_dir, env: SimEnv):
pyrado.set_seed(0)
# Environments
env_real = deepcopy(env)
env_real = ActNormWrapper(env_real)
env_sim = ActNormWrapper(env)
randomizer = DomainRandomizer(
NormalDomainParam(name="mass_rot_pole",
mean=0.0,
std=1e6,
clip_lo=1e-3),
NormalDomainParam(name="mass_pend_pole",
mean=0.0,
std=1e6,
clip_lo=1e-3),
NormalDomainParam(name="length_rot_pole",
mean=0.0,
std=1e6,
clip_lo=1e-3),
NormalDomainParam(name="length_pend_pole",
mean=0.0,
std=1e6,
clip_lo=1e-3),
)
env_sim = DomainRandWrapperLive(env_sim, randomizer)
dp_map = {
0: ("mass_rot_pole", "mean"),
1: ("mass_rot_pole", "std"),
2: ("mass_pend_pole", "mean"),
3: ("mass_pend_pole", "std"),
4: ("length_rot_pole", "mean"),
5: ("length_rot_pole", "std"),
6: ("length_pend_pole", "mean"),
7: ("length_pend_pole", "std"),
}
trafo_mask = [True] * 8
env_sim = MetaDomainRandWrapper(env_sim, dp_map)
# Subroutine for policy improvement
behav_policy_hparam = dict(hidden_sizes=[16, 16], hidden_nonlin=to.tanh)
behav_policy = FNNPolicy(spec=env_sim.spec, **behav_policy_hparam)
vfcn_hparam = dict(hidden_sizes=[16, 16], hidden_nonlin=to.relu)
vfcn = FNNPolicy(spec=EnvSpec(env_sim.obs_space, ValueFunctionSpace),
**vfcn_hparam)
critic_hparam = dict(
gamma=0.99,
lamda=0.98,
num_epoch=2,
batch_size=128,
standardize_adv=True,
lr=8e-4,
max_grad_norm=5.0,
)
critic = GAE(vfcn, **critic_hparam)
subrtn_policy_hparam = dict(
max_iter=2,
eps_clip=0.13,
min_steps=4 * env_sim.max_steps,
num_epoch=3,
batch_size=128,
std_init=0.75,
lr=3e-04,
max_grad_norm=1.0,
num_workers=1,
)
subrtn_policy = PPO(ex_dir, env_sim, behav_policy, critic,
**subrtn_policy_hparam)
prior = DomainRandomizer(
NormalDomainParam(name="mass_rot_pole", mean=0.095, std=0.095 / 10),
NormalDomainParam(name="mass_pend_pole", mean=0.024, std=0.024 / 10),
NormalDomainParam(name="length_rot_pole", mean=0.085, std=0.085 / 10),
NormalDomainParam(name="length_pend_pole", mean=0.129, std=0.129 / 10),
)
ddp_policy_hparam = dict(mapping=dp_map,
trafo_mask=trafo_mask,
scale_params=True)
ddp_policy = DomainDistrParamPolicy(prior=prior, **ddp_policy_hparam)
subsubrtn_distr_hparam = dict(
max_iter=2,
pop_size=10,
num_init_states_per_domain=1,
num_is_samples=8,
expl_std_init=1e-2,
expl_std_min=1e-5,
extra_expl_std_init=1e-2,
extra_expl_decay_iter=5,
num_workers=1,
)
subsubrtn_distr = CEM(ex_dir, env_sim, ddp_policy,
**subsubrtn_distr_hparam)
subrtn_distr_hparam = dict(
metric=None,
obs_dim_weight=[1, 1, 1, 1, 10, 10],
num_rollouts_per_distr=3,
num_workers=1,
)
subrtn_distr = SysIdViaEpisodicRL(subsubrtn_distr,
behavior_policy=behav_policy,
**subrtn_distr_hparam)
# Algorithm
algo_hparam = dict(
max_iter=1,
num_eval_rollouts=5,
warmstart=True,
)
algo = SimOpt(ex_dir, env_sim, env_real, subrtn_policy, subrtn_distr,
**algo_hparam)
algo.train()
assert algo.curr_iter == algo.max_iter
def test_sysidasrl_reps(ex_dir, env: SimEnv, num_eval_rollouts: int):
pyrado.set_seed(0)
def eval_ddp_policy(rollouts_real):
init_states_real = np.array([ro.states[0, :] for ro in rollouts_real])
rollouts_sim = []
for i, _ in enumerate(range(num_eval_rollouts)):
rollouts_sim.append(
rollout(env_sim,
behavior_policy,
eval=True,
reset_kwargs=dict(init_state=init_states_real[i, :])))
# Clip the rollouts rollouts yielding two lists of pairwise equally long rollouts
ros_real_tr, ros_sim_tr = algo.truncate_rollouts(rollouts_real,
rollouts_sim,
replicate=False)
assert len(ros_real_tr) == len(ros_sim_tr)
assert all([
np.allclose(r.states[0, :], s.states[0, :])
for r, s in zip(ros_real_tr, ros_sim_tr)
])
# Return the average the loss
losses = [
algo.loss_fcn(ro_r, ro_s)
for ro_r, ro_s in zip(ros_real_tr, ros_sim_tr)
]
return float(np.mean(np.asarray(losses)))
# Environments
env_real = deepcopy(env)
env_real.domain_param = dict(ang_offset=-2 * np.pi / 180)
env_sim = deepcopy(env)
randomizer = DomainRandomizer(
UniformDomainParam(name="ang_offset", mean=0, halfspan=1e-6), )
env_sim = DomainRandWrapperLive(env_sim, randomizer)
dp_map = {0: ("ang_offset", "mean"), 1: ("ang_offset", "halfspan")}
env_sim = MetaDomainRandWrapper(env_sim, dp_map)
assert env_real is not env_sim
# Policies (the behavioral policy needs to be deterministic)
behavior_policy = LinearPolicy(env_sim.spec,
feats=FeatureStack(identity_feat))
prior = DomainRandomizer(
UniformDomainParam(name="ang_offset",
mean=1 * np.pi / 180,
halfspan=1 * np.pi / 180), )
ddp_policy = DomainDistrParamPolicy(mapping=dp_map,
trafo_mask=[False, True],
prior=prior)
# Subroutine
subrtn_hparam = dict(
max_iter=2,
eps=1.0,
pop_size=100,
num_init_states_per_domain=1,
expl_std_init=5e-2,
expl_std_min=1e-4,
num_workers=1,
)
subrtn = REPS(ex_dir, env_sim, ddp_policy, **subrtn_hparam)
algo_hparam = dict(metric=None,
obs_dim_weight=np.ones(env_sim.obs_space.shape),
num_rollouts_per_distr=5,
num_workers=1)
algo = SysIdViaEpisodicRL(subrtn, behavior_policy, **algo_hparam)
rollouts_real_tst = []
for _ in range(num_eval_rollouts):
rollouts_real_tst.append(rollout(env_real, behavior_policy, eval=True))
loss_pre = eval_ddp_policy(rollouts_real_tst)
# Mimic training
while algo.curr_iter < algo.max_iter and not algo.stopping_criterion_met():
algo.logger.add_value(algo.iteration_key, algo.curr_iter)
# Creat fake real-world data
rollouts_real = []
for _ in range(num_eval_rollouts):
rollouts_real.append(rollout(env_real, behavior_policy, eval=True))
algo.step(snapshot_mode="latest",
meta_info=dict(rollouts_real=rollouts_real))
algo.logger.record_step()
algo._curr_iter += 1
loss_post = eval_ddp_policy(rollouts_real_tst)
assert loss_post <= loss_pre # don't have to be better every step
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)
# Environments
env_hparams = dict(dt=1 / 100., max_steps=600)
env_real = QQubeSwingUpSim(**env_hparams)
env_real.domain_param = dict(
Mr=0.095 * 0.9, # 0.095*0.9 = 0.0855
Mp=0.024 * 1.1, # 0.024*1.1 = 0.0264
Lr=0.085 * 0.9, # 0.085*0.9 = 0.0765
Lp=0.129 * 1.1, # 0.129*1.1 = 0.1419
)
env_sim = QQubeSwingUpSim(**env_hparams)
randomizer = DomainRandomizer(
NormalDomainParam(name='Mr', mean=0., std=1e6, clip_lo=1e-3),
NormalDomainParam(name='Mp', mean=0., std=1e6, clip_lo=1e-3),
NormalDomainParam(name='Lr', mean=0., std=1e6, clip_lo=1e-3),
NormalDomainParam(name='Lp', mean=0., std=1e6, clip_lo=1e-3),
)
env_sim = DomainRandWrapperLive(env_sim, randomizer)
dp_map = {
0: ('Mr', 'mean'),
1: ('Mr', 'std'),
2: ('Mp', 'mean'),
3: ('Mp', 'std'),
4: ('Lr', 'mean'),
5: ('Lr', 'std'),
6: ('Lp', 'mean'),
7: ('Lp', 'std')
}
trafo_mask = [True] * 8
env_sim = MetaDomainRandWrapper(env_sim, dp_map)
# Subroutine for policy improvement
behav_policy_hparam = dict(hidden_sizes=[64, 64], hidden_nonlin=to.tanh)
behav_policy = FNNPolicy(spec=env_sim.spec, **behav_policy_hparam)
vfcn_hparam = dict(hidden_sizes=[64, 64], hidden_nonlin=to.tanh)
vfcn = FNNPolicy(spec=EnvSpec(env_sim.obs_space, ValueFunctionSpace),
**vfcn_hparam)
critic_hparam = dict(
gamma=0.9885,
lamda=0.9648,
num_epoch=2,
batch_size=500,
standardize_adv=False,
lr=5.792e-4,
max_grad_norm=1.,
)
critic = GAE(vfcn, **critic_hparam)
subrtn_policy_hparam = dict(
max_iter=200,
min_steps=3 * 23 * env_sim.max_steps,
num_epoch=7,
eps_clip=0.0744,
batch_size=500,
std_init=0.9074,
lr=3.446e-04,
max_grad_norm=1.,
num_workers=1,
)
subrtn_policy = PPO(study_dir, env_sim, behav_policy, critic,
**subrtn_policy_hparam)
# Subroutine for system identification
prior_std_denom = trial.suggest_uniform('prior_std_denom', 5, 20)
prior = DomainRandomizer(
NormalDomainParam(name='Mr', mean=0.095, std=0.095 / prior_std_denom),
NormalDomainParam(name='Mp', mean=0.024, std=0.024 / prior_std_denom),
NormalDomainParam(name='Lr', mean=0.085, std=0.085 / prior_std_denom),
NormalDomainParam(name='Lp', mean=0.129, std=0.129 / prior_std_denom),
)
ddp_policy = DomainDistrParamPolicy(
mapping=dp_map,
trafo_mask=trafo_mask,
prior=prior,
scale_params=trial.suggest_categorical('ddp_policy_scale_params',
[True, False]),
)
subsubrtn_distr_hparam = dict(
max_iter=trial.suggest_categorical('subsubrtn_distr_max_iter', [20]),
pop_size=trial.suggest_int('pop_size', 50, 500),
num_rollouts=1,
num_is_samples=trial.suggest_int('num_is_samples', 5, 20),
expl_std_init=trial.suggest_loguniform('expl_std_init', 1e-3, 1e-1),
expl_std_min=trial.suggest_categorical('expl_std_min', [1e-4]),
extra_expl_std_init=trial.suggest_loguniform('expl_std_init', 1e-3,
1e-1),
extra_expl_decay_iter=trial.suggest_int('extra_expl_decay_iter', 0,
10),
num_workers=1,
)
csv_logger = create_csv_step_logger(
osp.join(study_dir, f'trial_{trial.number}'))
subsubrtn_distr = CEM(study_dir,
env_sim,
ddp_policy,
**subsubrtn_distr_hparam,
logger=csv_logger)
obs_vel_weight = trial.suggest_loguniform('obs_vel_weight', 1, 100)
subrtn_distr_hparam = dict(
metric=None,
obs_dim_weight=[1, 1, 1, 1, obs_vel_weight, obs_vel_weight],
num_rollouts_per_distr=trial.suggest_int('num_rollouts_per_distr', 20,
100),
num_workers=1,
)
subrtn_distr = SysIdViaEpisodicRL(subsubrtn_distr, behav_policy,
**subrtn_distr_hparam)
# Algorithm
algo_hparam = dict(
max_iter=trial.suggest_categorical('algo_max_iter', [10]),
num_eval_rollouts=trial.suggest_categorical('algo_num_eval_rollouts',
[5]),
warmstart=trial.suggest_categorical('algo_warmstart', [True]),
thold_succ_subrtn=trial.suggest_categorical('algo_thold_succ_subrtn',
[50]),
subrtn_snapshot_mode='latest',
)
algo = SimOpt(study_dir,
env_sim,
env_real,
subrtn_policy,
subrtn_distr,
**algo_hparam,
logger=csv_logger)
# Jeeeha
algo.train(seed=args.seed)
# Evaluate
min_rollouts = 1000
sampler = ParallelRolloutSampler(
env_real, algo.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