def test_multi_output(self):
Q5 = MLPValueFunction(
input_shapes=(
self.env.observation_space.shape,
self.env.action_space.shape,
),
output_shape=(5, ),
hidden_layer_sizes=self.hidden_layer_sizes,
name="Q5",
)
observation1_np = self.env.reset()
observation2_np = self.env.step(self.env.action_space.sample())[0]
action1_np = self.env.action_space.sample()
action2_np = self.env.action_space.sample()
observations_np = np.stack(
(observation1_np, observation2_np)).astype(np.float32)
actions_np = np.stack((action1_np, action2_np))
conditions = [observations_np, actions_np]
q_values_np = Q5.get_values_np(conditions)
q_values = Q5.get_values(conditions)
self.assertEqual(q_values_np.shape, (2, 5))
self.assertEqual(q_values.shape, (2, 5))
def get_pr2_agent(env,
agent_id,
hidden_layer_sizes,
max_replay_buffer_size,
policy_type="deter"):
observation_space = env.env_specs.observation_space[agent_id]
action_space = env.env_specs.action_space[agent_id]
opponent_action_shape = (
env.env_specs.action_space.opponent_flat_dim(agent_id), )
print(opponent_action_shape, "opponent_action_shape")
if policy_type == "dete":
policy_fn = DeterministicMLPPolicy
exploration_strategy = OUExploration(action_space)
elif policy_type == "gumble":
policy_fn = RelaxedSoftmaxMLPPolicy
exploration_strategy = None
return PR2Agent(
env_specs=env.env_specs,
policy=policy_fn(
input_shapes=(observation_space.shape, ),
output_shape=action_space.shape,
hidden_layer_sizes=hidden_layer_sizes,
name="policy_agent_{}".format(agent_id),
),
qf=MLPValueFunction(
input_shapes=(
observation_space.shape,
action_space.shape,
opponent_action_shape,
),
output_shape=(1, ),
hidden_layer_sizes=hidden_layer_sizes,
name="qf_agent_{}".format(agent_id),
),
ind_qf=MLPValueFunction(
input_shapes=(observation_space.shape, action_space.shape),
output_shape=(1, ),
hidden_layer_sizes=hidden_layer_sizes,
name="ind_qf_agent_{}".format(agent_id),
),
replay_buffer=IndexedReplayBuffer(
observation_dim=observation_space.shape[0],
action_dim=action_space.shape[0],
max_replay_buffer_size=max_replay_buffer_size,
opponent_action_dim=opponent_action_shape[0],
),
opponent_policy=policy_fn(
input_shapes=(observation_space.shape, action_space.shape),
output_shape=opponent_action_shape,
hidden_layer_sizes=hidden_layer_sizes,
name="opponent_policy_agent_{}".format(agent_id),
),
exploration_strategy=exploration_strategy,
gradient_clipping=10.0,
agent_id=agent_id,
)
def setUp(self):
self.env = gym.envs.make('MountainCarContinuous-v0')
self.hidden_layer_sizes = (128, 128)
self.Q = MLPValueFunction(
input_shapes=(self.env.observation_space.shape, self.env.action_space.shape),
output_shape=(1,),
hidden_layer_sizes=self.hidden_layer_sizes,
name='Q'
)
self.V = MLPValueFunction(
input_shapes=(self.env.observation_space.shape,),
output_shape=(1,),
hidden_layer_sizes=self.hidden_layer_sizes,
name='V'
)
def get_ddpgtom_agent(env, agent_id, hidden_layer_sizes,
max_replay_buffer_size):
observation_space = env.env_specs.observation_space[agent_id]
action_space = env.env_specs.action_space[agent_id]
return DDPGToMAgent(
env_specs=env.env_specs,
policy=DeterministicMLPPolicy(input_shapes=(observation_space.shape, (
env.env_specs.action_space.opponent_flat_dim(agent_id), )),
output_shape=action_space.shape,
hidden_layer_sizes=hidden_layer_sizes,
name='policy_agent_{}'.format(agent_id)),
qf=MLPValueFunction(
input_shapes=(observation_space.shape,
(env.env_specs.action_space.flat_dim, )),
output_shape=(1, ),
hidden_layer_sizes=hidden_layer_sizes,
name='qf_agent_{}'.format(agent_id)),
opponent_policy=DeterministicMLPPolicy(
input_shapes=(observation_space.shape, ),
output_shape=(
env.env_specs.action_space.opponent_flat_dim(agent_id), ),
hidden_layer_sizes=hidden_layer_sizes,
name='opponent_policy_agent_{}'.format(agent_id)),
replay_buffer=IndexedReplayBuffer(
observation_dim=observation_space.shape[0],
action_dim=action_space.shape[0],
opponent_action_dim=env.env_specs.action_space.opponent_flat_dim(
agent_id),
max_replay_buffer_size=max_replay_buffer_size),
exploration_strategy=OUExploration(action_space),
gradient_clipping=10.,
agent_id=agent_id,
)
def get_ddpg_agent(env,
agent_id,
hidden_layer_sizes,
max_replay_buffer_size,
policy_type='dete'):
observation_space = env.env_specs.observation_space[agent_id]
action_space = env.env_specs.action_space[agent_id]
if policy_type == 'dete':
policy_fn = DeterministicMLPPolicy
exploration_strategy = OUExploration(action_space)
elif policy_type == 'gumble':
policy_fn = RelaxedSoftmaxMLPPolicy
exploration_strategy = None
return DDPGAgent(
env_specs=env.env_specs,
policy=policy_fn(input_shapes=(observation_space.shape, ),
output_shape=action_space.shape,
hidden_layer_sizes=hidden_layer_sizes,
name='policy_agent_{}'.format(agent_id)),
qf=MLPValueFunction(input_shapes=(observation_space.shape,
action_space.shape),
output_shape=(1, ),
hidden_layer_sizes=hidden_layer_sizes,
name='qf_agent_{}'.format(agent_id)),
replay_buffer=IndexedReplayBuffer(
observation_dim=observation_space.shape[0],
action_dim=action_space.shape[0],
max_replay_buffer_size=max_replay_buffer_size),
exploration_strategy=exploration_strategy,
gradient_clipping=10.,
agent_id=agent_id,
)
def get_rommeo_agent(
env,
agent_id,
hidden_layer_sizes,
max_replay_buffer_size,
policy_type="gaussian",
uniform=False,
custom_b=False,
bi=1.0,
bj=1.0,
):
observation_space = env.env_specs.observation_space[agent_id]
action_space = env.env_specs.action_space[agent_id]
opponent_action_shape = (
env.env_specs.action_space.opponent_flat_dim(agent_id), )
if policy_type == "gaussian":
policy_fn = GaussianMLPPolicy
elif policy_type == "gumble":
policy_fn = RelaxedSoftmaxMLPPolicy
return ROMMEOAgent(
env_specs=env.env_specs,
policy=policy_fn(
input_shapes=(observation_space.shape, opponent_action_shape),
output_shape=action_space.shape,
hidden_layer_sizes=hidden_layer_sizes,
name="policy_agent_{}".format(agent_id),
repara=True,
# smoothing_coefficient=0.5
),
qf=MLPValueFunction(
input_shapes=(
observation_space.shape,
action_space.shape,
opponent_action_shape,
),
output_shape=(1, ),
hidden_layer_sizes=hidden_layer_sizes,
name="qf_agent_{}".format(agent_id),
),
replay_buffer=IndexedReplayBuffer(
observation_dim=observation_space.shape[0],
action_dim=action_space.shape[0],
max_replay_buffer_size=max_replay_buffer_size,
opponent_action_dim=opponent_action_shape[0],
),
opponent_policy=policy_fn(
input_shapes=(observation_space.shape, ),
output_shape=opponent_action_shape,
hidden_layer_sizes=hidden_layer_sizes,
name="opponent_policy_agent_{}".format(agent_id),
repara=True,
),
gradient_clipping=10,
agent_id=agent_id,
name="ROMMEO_{}".format(agent_id),
uniform=uniform,
custom_b=custom_b,
bi=bi,
bj=bj,
)
def get_pr2k_soft_agent(env,
agent_id,
hidden_layer_sizes,
max_replay_buffer_size,
k=2,
mu=0):
observation_space = env.env_specs.observation_space[agent_id]
action_space = env.env_specs.action_space[agent_id]
opponent_action_shape = (
env.env_specs.action_space.opponent_flat_dim(agent_id), )
print(opponent_action_shape, "opponent_action_shape")
return PR2KSoftAgent(
env_specs=env.env_specs,
main_policy=GaussianMLPPolicy(
input_shapes=(observation_space.shape, opponent_action_shape),
output_shape=action_space.shape,
hidden_layer_sizes=hidden_layer_sizes,
name="policy_agent_{}".format(agent_id),
),
opponent_policy=GaussianMLPPolicy(
input_shapes=(observation_space.shape, action_space.shape),
output_shape=opponent_action_shape,
hidden_layer_sizes=hidden_layer_sizes,
name="opponent_policy_agent_{}".format(agent_id),
),
prior_policy=GaussianMLPPolicy(
input_shapes=(observation_space.shape, ),
output_shape=action_space.shape,
hidden_layer_sizes=hidden_layer_sizes,
name="prior_policy_agent_{}".format(agent_id),
),
opponent_prior_policy=GaussianMLPPolicy(
input_shapes=(observation_space.shape, ),
output_shape=opponent_action_shape,
hidden_layer_sizes=hidden_layer_sizes,
name="opponent_prior_policy_agent_{}".format(agent_id),
),
qf=MLPValueFunction(
input_shapes=(
observation_space.shape,
action_space.shape,
opponent_action_shape,
),
output_shape=(1, ),
hidden_layer_sizes=hidden_layer_sizes,
name="qf_agent_{}".format(agent_id),
),
replay_buffer=IndexedReplayBuffer(
observation_dim=observation_space.shape[0],
action_dim=action_space.shape[0],
max_replay_buffer_size=max_replay_buffer_size,
opponent_action_dim=opponent_action_shape[0],
),
k=k,
mu=mu,
gradient_clipping=10.0,
agent_id=agent_id,
)
def get_sac_agent(env,
hidden_layer_sizes,
max_replay_buffer_size,
policy_type="gaussian"):
"""
SAC agent for single player learning.
"""
observation_space = env.env_specs.observation_space[0]
action_space = env.env_specs.action_space[0]
env_specs = env.env_specs
if policy_type == "gaussian":
policy_fn = GaussianMLPPolicy
elif policy_type == "gumble":
policy_fn = RelaxedSoftmaxMLPPolicy
# print('observation_space.shape', observation_space.shape)
return SACAgent(
env_specs=env_specs,
policy=policy_fn(
input_shapes=(observation_space.shape, ),
output_shape=action_space.shape,
hidden_layer_sizes=hidden_layer_sizes,
name="{}_policy".format(policy_type),
),
qfs=[
MLPValueFunction(
input_shapes=(observation_space.shape, action_space.shape),
output_shape=(1, ),
hidden_layer_sizes=hidden_layer_sizes,
name="qf_{}".format(qf_id),
) for qf_id in range(2)
],
vf=MLPValueFunction(
input_shapes=(observation_space.shape, ),
output_shape=(1, ),
hidden_layer_sizes=hidden_layer_sizes,
name="vf",
),
replay_buffer=IndexedReplayBuffer(
observation_dim=observation_space.shape[0],
action_dim=action_space.shape[0],
max_replay_buffer_size=max_replay_buffer_size,
),
)
def get_pr2_soft_agent(env,
agent_id,
hidden_layer_sizes,
max_replay_buffer_size,
policy_type="gaussian"):
observation_space = env.env_specs.observation_space[agent_id]
action_space = env.env_specs.action_space[agent_id]
opponent_action_shape = (
env.env_specs.action_space.opponent_flat_dim(agent_id), )
if policy_type == "gaussian":
policy_fn = GaussianMLPPolicy
elif policy_type == "gumble":
policy_fn = RelaxedSoftmaxMLPPolicy
return PR2SoftAgent(
env_specs=env.env_specs,
policy=policy_fn(
input_shapes=(observation_space.shape, ),
output_shape=action_space.shape,
hidden_layer_sizes=hidden_layer_sizes,
name="policy_agent_{}".format(agent_id),
),
qf=MLPValueFunction(
input_shapes=(
observation_space.shape,
action_space.shape,
opponent_action_shape,
),
output_shape=(1, ),
hidden_layer_sizes=hidden_layer_sizes,
name="qf_agent_{}".format(agent_id),
),
replay_buffer=IndexedReplayBuffer(
observation_dim=observation_space.shape[0],
action_dim=action_space.shape[0],
max_replay_buffer_size=max_replay_buffer_size,
opponent_action_dim=opponent_action_shape[0],
),
opponent_policy=policy_fn(
input_shapes=(observation_space.shape, action_space.shape),
output_shape=opponent_action_shape,
hidden_layer_sizes=hidden_layer_sizes,
name="opponent_policy_agent_{}".format(agent_id),
),
gradient_clipping=10.0,
agent_id=agent_id,
)
class ValueFunctionTest(tf.test.TestCase):
def setUp(self):
self.env = gym.envs.make('MountainCarContinuous-v0')
self.hidden_layer_sizes = (128, 128)
self.Q = MLPValueFunction(
input_shapes=(self.env.observation_space.shape, self.env.action_space.shape),
output_shape=(1,),
hidden_layer_sizes=self.hidden_layer_sizes,
name='Q'
)
self.V = MLPValueFunction(
input_shapes=(self.env.observation_space.shape,),
output_shape=(1,),
hidden_layer_sizes=self.hidden_layer_sizes,
name='V'
)
def test_multi_output(self):
Q5 = MLPValueFunction(
input_shapes=(self.env.observation_space.shape, self.env.action_space.shape),
output_shape=(5,),
hidden_layer_sizes=self.hidden_layer_sizes,
name='Q5'
)
observation1_np = self.env.reset()
observation2_np = self.env.step(self.env.action_space.sample())[0]
action1_np = self.env.action_space.sample()
action2_np = self.env.action_space.sample()
observations_np = np.stack((observation1_np, observation2_np)).astype(np.float32)
actions_np = np.stack((action1_np, action2_np))
conditions = [observations_np, actions_np]
q_values_np = Q5.get_values_np(conditions)
q_values = Q5.get_values(conditions)
self.assertEqual(q_values_np.shape, (2, 5))
self.assertEqual(q_values.shape, (2, 5))
def test_values(self):
observation1_np = self.env.reset()
observation2_np = self.env.step(self.env.action_space.sample())[0]
action1_np = self.env.action_space.sample()
action2_np = self.env.action_space.sample()
observations_np = np.stack((observation1_np, observation2_np)).astype(np.float32)
actions_np = np.stack((action1_np, action2_np))
conditions = [observations_np, actions_np]
q_values_np = self.Q.get_values_np(conditions)
q_values = self.Q.get_values(conditions)
v_values_np = self.V.get_values_np([observations_np])
v_values = self.V.get_values([observations_np])
self.assertEqual(q_values_np.shape, (2, 1))
self.assertEqual(q_values.shape, (2, 1))
self.assertEqual(v_values_np.shape, (2, 1))
self.assertEqual(v_values.shape, (2, 1))
def test_trainable_variables(self):
self.assertEqual(
len(self.Q.trainable_variables),
2 * (len(self.hidden_layer_sizes) + 1))
self.assertEqual(
len(self.V.trainable_variables),
2 * (len(self.hidden_layer_sizes) + 1))
def test_clone_target(self):
observation1_np = self.env.reset()
observation2_np = self.env.step(self.env.action_space.sample())[0]
observations_np = np.stack(
(observation1_np, observation2_np)
).astype(np.float32)
weights = self.V.get_weights()
values_np = self.V.get_values_np([observations_np])
target_name = '{}_{}'.format('target', self.V._name)
target_V = Serializable.clone(self.V, name=target_name)
weights_2 = target_V.get_weights()
self.assertEqual(target_V._name, target_name)
self.assertIsNot(weights, weights_2)
for weight, weight_2 in zip(weights, weights_2):
np.testing.assert_array_equal(weight.shape, weight_2.shape)
np.testing.assert_equal(
values_np.shape, target_V.get_values_np([observations_np]).shape)
def test_serialize_deserialize(self):
observation1_np = self.env.reset()
observation2_np = self.env.step(self.env.action_space.sample())[0]
observations_np = np.stack(
(observation1_np, observation2_np)
).astype(np.float32)
weights = self.V.get_weights()
values_np = self.V.get_values_np([observations_np])
serialized = pickle.dumps(self.V)
deserialized = pickle.loads(serialized)
weights_2 = deserialized.get_weights()
for weight, weight_2 in zip(weights, weights_2):
np.testing.assert_array_equal(weight, weight_2)
np.testing.assert_equal(
values_np.shape, deserialized.get_values_np([observations_np]).shape)