# Set seed if desired
pyrado.set_seed(args.seed, verbose=True)
# Environment
env_hparams = dict(dt=1 / 100.0, max_steps=600)
env = QQubeSwingUpSim(**env_hparams)
# env = ObsVelFiltWrapper(env, idcs_pos=["theta", "alpha"], idcs_vel=["theta_dot", "alpha_dot"])
env = ActNormWrapper(env)
# Policy
if args.mode is None or args.mode.lower() == FNNPolicy.name:
policy_hparam = dict(hidden_sizes=[64, 64], hidden_nonlin=to.tanh)
policy = FNNPolicy(spec=env.spec, **policy_hparam)
elif args.mode.lower() == GRUPolicy.name:
policy_hparam = dict(hidden_size=32, num_recurrent_layers=1)
policy = GRUPolicy(spec=env.spec, **policy_hparam)
# Critic
if isinstance(policy, FNNPolicy):
vfcn_hparam = dict(hidden_sizes=[32, 32], hidden_nonlin=to.relu) # FNN
vfcn = FNNPolicy(spec=EnvSpec(env.obs_space, ValueFunctionSpace), **vfcn_hparam)
elif isinstance(policy, GRUPolicy):
vfcn_hparam = dict(hidden_size=32, num_recurrent_layers=1) # LSTM & GRU
vfcn = GRUPolicy(spec=EnvSpec(env.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,
def gru_policy(env: Env):
return GRUPolicy(env.spec, hidden_size=8, num_recurrent_layers=2)
def gru_policy_cuda(env: Env):
return GRUPolicy(env.spec, hidden_size=8, num_recurrent_layers=2, use_cuda=True)
# Set seed if desired
pyrado.set_seed(args.seed, verbose=True)
# Environment
env_hparams = dict(dt=1 / 250.0, max_steps=12 * 250, long=False)
env = QCartPoleSwingUpSim(**env_hparams)
env = ActNormWrapper(env)
# Policy
policy_hparam = dict(
hidden_size=32,
num_recurrent_layers=1,
# init_param_kwargs=dict(t_max=50)
)
# policy = LSTMPolicy(spec=env.spec, **policy_hparam)
policy = GRUPolicy(spec=env.spec, **policy_hparam)
# Algorithm
algo_hparam = dict(
max_iter=5000,
pop_size=50,
num_init_states_per_domain=6,
eta_mean=2.0,
eta_std=None,
expl_std_init=0.5,
symm_sampling=False,
transform_returns=True,
num_workers=10,
)
algo = NES(ex_dir, env, policy, **algo_hparam)