'states': states,
'actions': actions,
'transform_to_internal_state': lambda x: (math.atan2(x[1], x[0]), x[
2]), # use when the obsevation and state space do not exactly match
'render': False
}
# sampling_params = {'sampling_type': 'behavioral',
# 'transform_to_internal_state': lambda x: (math.atan2(x[1], x[0]), x[2]),
# 'initial_state': np.array([1., 0., 0.], dtype=NP_DTYPE), # theta=0., theta_dot=0,
# 'policy': policy_gmm,
# 'n_samples': 1000, # n_trajectories: 2
# 'max_ep_transitions': 500
# }
dataset = mdp.get_samples(**sampling_params)
dataset.update_dataset_internal()
# Compute Kernel Bandwidths (state, action, state next)
s_band_factor = [1., 1., 1.] # Pendulum uniform
# s_band_factor = [10., 10., 1.] # Pendulum behavioral
s_n_band_factor = s_band_factor
a_band_factor = [50.] # Pendulum uniform
# a_band_factor = [10.] # Pendulum behavioral
dataset.kde_bandwidths_internal(s_band_factor=s_band_factor,
a_band_factor=a_band_factor,
s_n_band_factor=s_n_band_factor)
##########################################################################################
# Define the Policy Network
def test_nopg_d_cartpole_behavioral(self):
##########################################################################################
# Create the Environment (MDP)
env = gym.make('CartpoleStabShort-v1')
mdp = MDP(env)
##########################################################################################
# Gather an Off-Policy Dataset
sampling_params = {
'sampling_type': 'behavioral',
'policy':
lambda x: np.random.uniform(low=-5., high=5., size=(1, )),
'n_samples': 200, # n_trajectories: 2
}
dataset = mdp.get_samples(**sampling_params)
dataset.update_dataset_internal()
# Compute Kernel Bandwidths (state, action, state next)
s_band_factor = [1., 1., 1., 1., 1.]
s_n_band_factor = s_band_factor
a_band_factor = [20.]
dataset.kde_bandwidths_internal(s_band_factor=s_band_factor,
a_band_factor=a_band_factor,
s_n_band_factor=s_n_band_factor)
##########################################################################################
# Define the Policy Network
policy_params = {
'policy_class': 'deterministic',
'neurons':
[mdp.s_dim, 10,
mdp.a_dim], # [state_dim, hidden1, ... , hiddenN, action_dim]
'activations':
[nn.functional.relu], # one activation function per hidden layer
'f_out': [lambda x: 5.0 * torch.tanh(x)],
'device': DEVICE
}
policy = Policy(**policy_params).to(device=DEVICE, dtype=TORCH_DTYPE)
##########################################################################################
# Optimize the policy with NOPG
nopg_params = {
'initial_states':
np.array([env.reset()
for _ in range(5)]), # For multiple initial states
'gamma': 0.99
}
nopg = NOPG(dataset, policy, **nopg_params)
n_policy_updates = 10
def optimizer(x):
return optim.Adam(x, lr=1e-2)
evaluation_params = {
'eval_mdp': mdp,
'eval_every_n': 200,
'eval_n_episodes': 1,
'eval_render': False
}
nopg.fit(n_policy_updates=n_policy_updates,
optimizer=optimizer,
**evaluation_params)
def test_nopg_s_pendulum_behavioral_sample_P(self):
##########################################################################################
# Create the Environment (MDP)
env = gym.make('Pendulum-v1')
mdp = MDP(env)
##########################################################################################
# Gather an Off-Policy Dataset
sampling_params = {
'sampling_type': 'behavioral',
'transform_to_internal_state': lambda x:
(math.atan2(x[1], x[0]), x[2]),
'initial_state':
np.array([1., 0., 0.], dtype=NP_DTYPE), # theta=0., theta_dot=0,
'policy': policy_gmm,
'n_samples': 100, # n_trajectories: 2
'max_ep_transitions': 500
}
dataset = mdp.get_samples(**sampling_params)
dataset.update_dataset_internal()
# Compute Kernel Bandwidths (state, action, state next)
s_band_factor = [10., 10., 1.] # Pendulum behavioral
s_n_band_factor = s_band_factor
a_band_factor = [10.] # Pendulum behavioral
dataset.kde_bandwidths_internal(s_band_factor=s_band_factor,
a_band_factor=a_band_factor,
s_n_band_factor=s_n_band_factor)
##########################################################################################
# Define the Policy Network
policy_params = {
'policy_class':
'stochastic',
'neurons':
[mdp.s_dim, 10,
mdp.a_dim], # [state_dim, hidden1, ... , hiddenN, action_dim]
'activations':
[nn.functional.relu], # one activation function per hidden layer
'f_out':
[lambda x: 2.0 * torch.tanh(x), lambda x: 2.0 * torch.sigmoid(x)],
'device':
DEVICE
}
policy = Policy(**policy_params).to(device=DEVICE, dtype=TORCH_DTYPE)
##########################################################################################
# Optimize the policy with NOPG
nopg_params = {
'initial_states': np.array([-1., 0., 0.]).reshape((-1, 3)),
'gamma': 0.97,
'MC_samples_P': 2,
'MC_samples_stochastic_policy': 2
}
nopg = NOPG(dataset, policy, **nopg_params)
n_policy_updates = 10
def optimizer(x):
return optim.Adam(x, lr=1e-2)
evaluation_params = {
'eval_mdp':
mdp,
'eval_every_n':
200,
'eval_n_episodes':
1,
'eval_initial_state':
np.array([-1., 0., 0.]), # or None
'eval_transform_to_internal_state':
lambda x: (math.atan2(x[1], x[0]), x[2]), # or None
'eval_render':
False
}
nopg.fit(n_policy_updates=n_policy_updates,
optimizer=optimizer,
**evaluation_params)
def test_nopg_d_pendulum_uniform_sparsify_P_kl_divergence(self):
##########################################################################################
# Create the Environment (MDP)
env = gym.make('Pendulum-v1')
mdp = MDP(env)
##########################################################################################
# Gather an Off-Policy Dataset
states = mdp.discretize_space(space='state',
levels=[10, 10]) # theta, theta_dot
actions = mdp.discretize_space(space='action', levels=[2])
sampling_params = {
'sampling_type': 'uniform',
'states': states,
'actions': actions,
'transform_to_internal_state': lambda x:
(math.atan2(x[1], x[0]), x[2]),
'render': False
}
dataset = mdp.get_samples(**sampling_params)
dataset.update_dataset_internal()
# Compute Kernel Bandwidths (state, action, state next)
s_band_factor = [1., 1., 1.] # Pendulum uniform
s_n_band_factor = s_band_factor
a_band_factor = [50.] # Pendulum uniform
dataset.kde_bandwidths_internal(s_band_factor=s_band_factor,
a_band_factor=a_band_factor,
s_n_band_factor=s_n_band_factor)
##########################################################################################
# Define the Policy Network
policy_params = {
'policy_class': 'deterministic',
'neurons':
[mdp.s_dim, 10,
mdp.a_dim], # [state_dim, hidden1, ... , hiddenN, action_dim]
'activations':
[nn.functional.relu], # one activation function per hidden layer
'f_out': [lambda x: 2.0 * torch.tanh(x)],
'device': DEVICE
}
policy = Policy(**policy_params).to(device=DEVICE, dtype=TORCH_DTYPE)
##########################################################################################
# Optimize the policy with NOPG
nopg_params = {
'initial_states': np.array([-1., 0., 0.]).reshape((-1, 3)),
'gamma': 0.97,
'sparsify_P': {
'kl_max': 0.001,
'kl_interval_k': 10,
'kl_repeat_every_n_iterations': 3
}
}
nopg = NOPG(dataset, policy, **nopg_params)
n_policy_updates = 10
def optimizer(x):
return optim.Adam(x, lr=1e-2)
evaluation_params = {
'eval_mdp':
mdp,
'eval_every_n':
200,
'eval_n_episodes':
1,
'eval_initial_state':
np.array([-1., 0., 0.]), # or None
'eval_transform_to_internal_state':
lambda x: (math.atan2(x[1], x[0]), x[2]), # or None
'eval_render':
False
}
nopg.fit(n_policy_updates=n_policy_updates,
optimizer=optimizer,
**evaluation_params)
def test_nopg_d_pendulum_uniform(self):
##########################################################################################
# Create the Environment (MDP)
env = gym.make('Pendulum-v1')
mdp = MDP(env)
##########################################################################################
# Gather an Off-Policy Dataset
states = mdp.discretize_space(space='state',
levels=[10, 10]) # theta, theta_dot
actions = mdp.discretize_space(space='action', levels=[2])
sampling_params = {
'sampling_type': 'uniform',
'states': states,
'actions': actions,
'transform_to_internal_state': lambda x:
(math.atan2(x[1], x[0]), x[2]),
'render': False
}
# sampling_params = {'sampling_type': 'behavioral',
# 'transform_to_internal_state': lambda x: (math.atan2(x[1], x[0]), x[2]),
# 'initial_state': np.array([1., 0., 0.], dtype=NP_DTYPE), # theta=0., theta_dot=0,
# 'policy': policy_gmm,
# 'n_samples': 1000, # n_trajectories: 2
# 'max_ep_transitions': 500
# }
dataset = mdp.get_samples(**sampling_params)
dataset.update_dataset_internal()
# Compute Kernel Bandwidths (state, action, state next)
s_band_factor = [1., 1., 1.] # Pendulum uniform
# s_band_factor = [10., 10., 1.] # Pendulum behavioral
s_n_band_factor = s_band_factor
a_band_factor = [50.] # Pendulum uniform
# a_band_factor = [10.] # Pendulum behavioral
dataset.kde_bandwidths_internal(s_band_factor=s_band_factor,
a_band_factor=a_band_factor,
s_n_band_factor=s_n_band_factor)
##########################################################################################
# Define the Policy Network
policy_params = {
'policy_class': 'deterministic',
'neurons':
[mdp.s_dim, 10,
mdp.a_dim], # [state_dim, hidden1, ... , hiddenN, action_dim]
'activations':
[nn.functional.relu], # one activation function per hidden layer
'f_out': [lambda x: 2.0 * torch.tanh(x)],
'device': DEVICE
# NOTE: for a stochastic policy f_out is a list with 2 entries (mean and diagonal covariance)
# 'type': 'stochastic',
# 'f_out': [lambda x: 2.0 * torch.tanh(x), lambda x: 2.0 * torch.sigmoid(x)]
}
policy = Policy(**policy_params).to(device=DEVICE, dtype=TORCH_DTYPE)
##########################################################################################
# Optimize the policy with NOPG
nopg_params = {
'initial_states': np.array([-1., 0., 0.]).reshape((-1, 3)),
# 'initial_states': np.array([env.reset() for _ in range(20)]), # For multiple initial states
'gamma': 0.97
# 'MC_samples_stochastic_policy': 15
# 'MC_samples_P': 15
# 'sparsify_P': {'P_sparse_k': 10 }
# 'sparsify_P': {'kl_max': 0.001, 'kl_interval_k': 20, 'kl_repeat_every_n_iterations': 200}
}
nopg = NOPG(dataset, policy, **nopg_params)
n_policy_updates = 10
def optimizer(x):
return optim.Adam(x, lr=1e-2)
evaluation_params = {
'eval_mdp':
mdp,
'eval_every_n':
200,
'eval_n_episodes':
1,
'eval_initial_state':
np.array([-1., 0., 0.]), # or None
'eval_transform_to_internal_state':
lambda x: (math.atan2(x[1], x[0]), x[2]), # or None
'eval_render':
False
}
nopg.fit(n_policy_updates=n_policy_updates,
optimizer=optimizer,
**evaluation_params)
