ref_energy=init_loc[0],
energy_gain=init_loc[1],
# energy_th_gain=0.3, # This parameter is fixed.
energy_th_gain=init_loc[2], # This parameter is fixed.
acc_max=5.,
alpha_max_pd_enable=10.,
pd_gains=to.tensor(
[init_loc[3], init_loc[4], init_loc[5], init_loc[6]],
dtype=to.float64))
policy = QQubeSwingUpAndBalanceCtrl(env.spec, **policy_hparam)
# Set the policy parameters to the initial ones...
# policy.param_values = to.tensor(init_loc)
# Sample a policy from the final search distribution
policy.param_values = to.tensor(dist.sample([1]).view(-1))
# Algorithm
algo_hparam = dict(max_iter=50,
pop_size=1,
num_rollouts=1,
expl_std_init=1.0,
expl_std_min=0.0,
num_sampler_envs=1,
n_mc_samples_gradient=1,
coupling=True,
lr=1e-1,
optim='Adam',
real_env=True)
algo = EMVD(ex_dir, env, policy, dist, **algo_hparam)
# Policy
policy_hparam = dict(
ref_energy=init_loc[0],
energy_gain=init_loc[1],
# energy_th_gain=0.3, # This parameter is fixed.
energy_th_gain=init_loc[2], # This parameter is fixed.
acc_max=5.,
alpha_max_pd_enable=10.,
pd_gains=to.tensor([-1.7313308, 35.976177, -1.58682, 3.0102878],
dtype=to.float64))
policy = QQubeSwingUpAndBalanceCtrl(env.spec,
**policy_hparam,
only_swingup_control=True)
# Set the policy parameters to the initial ones...
policy.param_values = to.tensor(init_loc)
# Algorithm
algo_hparam = dict(max_iter=50,
pop_size=1,
num_rollouts=1,
expl_std_init=1.0,
expl_std_min=0.0,
num_sampler_envs=16,
n_mc_samples_gradient=1,
coupling=True,
lr=1e-1,
optim='Adam')
algo = EMVD(ex_dir, env, policy, dist, **algo_hparam)