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 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
lr_scheduler_hparam=dict(gamma=0.999),
)
subrtn_policy = PPO(ex_dir, env_sim, behav_policy, critic,
**subrtn_policy_hparam)
# Subroutine for system identification
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=False)
ddp_policy = DomainDistrParamPolicy(prior=prior, **ddp_policy_hparam)
subsubrtn_distr_hparam = dict(
max_iter=10,
pop_size=None,
eta_mean=1.0,
num_init_states_per_domain=1,
expl_std_init=5e-2,
expl_std_min=1e-4,
num_workers=num_workers,
)
subsubrtn_distr = NES(ex_dir, env_sim, ddp_policy,
**subsubrtn_distr_hparam)
subrtn_distr_hparam = dict(
metric=None,
obs_dim_weight=[1.0, 1.0, 1.0, 1.0, 2.0, 2.0],
num_rollouts_per_distr=len(dp_map) * 10,
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