def train(agent, environment, training_loops, steps_per_loop, output_path):
optimal_reward_fn = functools.partial(
env_util.compute_optimal_reward_with_classification_environment,
environment=environment,
)
optimal_action_fn = functools.partial(
env_util.compute_optimal_action_with_classification_environment,
environment=environment,
)
regret_metric = tf_bandit_metrics.RegretMetric(optimal_reward_fn)
cumulative_reward = CumulativeRewardMetric()
suboptimal_arms_metric = tf_bandit_metrics.SuboptimalArmsMetric(
optimal_action_fn)
metrics = [regret_metric, suboptimal_arms_metric, cumulative_reward]
from datetime import datetime
t1 = datetime.now()
trainer.train(
root_dir=output_path,
agent=agent,
environment=environment,
training_loops=training_loops,
steps_per_loop=steps_per_loop, # 452950//batch_size,
additional_metrics=metrics,
)
t2 = datetime.now()
print("Training time in minutes:")
print((t2 - t1).total_seconds() / 60)
def main(unused_argv):
tf.compat.v1.enable_v2_behavior() # The trainer only runs with V2 enabled.
means = [0.1, 0.2, 0.3, 0.45, 0.5]
env = bern_env.BernoulliPyEnvironment(means=means, batch_size=BATCH_SIZE)
environment = tf_py_environment.TFPyEnvironment(env)
def optimal_reward_fn(unused_observation):
return np.max(means)
def optimal_action_fn(unused_observation):
return np.int32(np.argmax(means))
if FLAGS.agent == 'BernTS':
agent = bern_ts_agent.BernoulliThompsonSamplingAgent(
time_step_spec=environment.time_step_spec(),
action_spec=environment.action_spec(),
dtype=tf.float64,
batch_size=BATCH_SIZE)
else:
raise ValueError('Only BernoulliTS is supported for now.')
regret_metric = tf_bandit_metrics.RegretMetric(optimal_reward_fn)
suboptimal_arms_metric = tf_bandit_metrics.SuboptimalArmsMetric(
optimal_action_fn)
trainer.train(root_dir=FLAGS.root_dir,
agent=agent,
environment=environment,
training_loops=TRAINING_LOOPS,
steps_per_loop=STEPS_PER_LOOP,
additional_metrics=[regret_metric, suboptimal_arms_metric],
save_policy=False)
def main(unused_argv):
tf.compat.v1.enable_v2_behavior() # The trainer only runs with V2 enabled.
data_path = FLAGS.data_path
if not data_path:
raise ValueError('Please specify the location of the data file.')
env = movielens_py_environment.MovieLensPyEnvironment(
data_path, RANK_K, BATCH_SIZE, num_movies=20)
environment = tf_py_environment.TFPyEnvironment(env)
optimal_reward_fn = functools.partial(
environment_utilities.compute_optimal_reward_with_movielens_environment,
environment=environment)
optimal_action_fn = functools.partial(
environment_utilities.compute_optimal_action_with_movielens_environment,
environment=environment)
if FLAGS.agent == 'LinUCB':
agent = lin_ucb_agent.LinearUCBAgent(
time_step_spec=environment.time_step_spec(),
action_spec=environment.action_spec(),
tikhonov_weight=0.001,
alpha=AGENT_ALPHA,
dtype=tf.float32)
elif FLAGS.agent == 'LinTS':
agent = lin_ts_agent.LinearThompsonSamplingAgent(
time_step_spec=environment.time_step_spec(),
action_spec=environment.action_spec(),
dtype=tf.float32)
elif FLAGS.agent == 'epsGreedy':
network = q_network.QNetwork(
input_tensor_spec=environment.time_step_spec().observation,
action_spec=environment.action_spec(),
fc_layer_params=LAYERS)
agent = eps_greedy_agent.NeuralEpsilonGreedyAgent(
time_step_spec=environment.time_step_spec(),
action_spec=environment.action_spec(),
reward_network=network,
optimizer=tf.compat.v1.train.AdamOptimizer(learning_rate=LR),
epsilon=EPSILON)
elif FLAGS.agent == 'DropoutTS':
agent = dropout_ts_agent.DropoutThompsonSamplingAgent(
time_step_spec=environment.time_step_spec(),
action_spec=environment.action_spec(),
dropout_rate=DROPOUT_RATE,
network_layers=LAYERS,
optimizer=tf.compat.v1.train.AdamOptimizer(learning_rate=LR))
regret_metric = tf_bandit_metrics.RegretMetric(optimal_reward_fn)
suboptimal_arms_metric = tf_bandit_metrics.SuboptimalArmsMetric(
optimal_action_fn)
trainer.train(
root_dir=FLAGS.root_dir,
agent=agent,
environment=environment,
training_loops=TRAINING_LOOPS,
steps_per_loop=STEPS_PER_LOOP,
additional_metrics=[regret_metric, suboptimal_arms_metric])
def main(unused_argv):
tf.compat.v1.enable_v2_behavior() # The trainer only runs with V2 enabled.
class LinearNormalReward(object):
def __init__(self, theta):
self.theta = theta
def __call__(self, x):
mu = np.dot(x, self.theta)
return np.random.normal(mu, 1)
def _global_context_sampling_fn():
return np.random.randint(-10, 10, [4]).astype(np.float32)
def _arm_context_sampling_fn():
return np.random.randint(-2, 3, [5]).astype(np.float32)
reward_fn = LinearNormalReward(HIDDEN_PARAM)
env = sspe.StationaryStochasticPerArmPyEnvironment(
_global_context_sampling_fn,
_arm_context_sampling_fn,
NUM_ACTIONS,
reward_fn,
batch_size=BATCH_SIZE)
environment = tf_py_environment.TFPyEnvironment(env)
obs_spec = environment.observation_spec()
if FLAGS.network == 'commontower':
network = (global_and_arm_feature_network.
create_feed_forward_common_tower_network(
obs_spec, (4, 3), (3, 4), (4, 2)))
elif FLAGS.network == 'dotproduct':
network = (global_and_arm_feature_network.
create_feed_forward_dot_product_network(
obs_spec, (4, 3, 6), (3, 4, 6)))
agent = neural_epsilon_greedy_agent.NeuralEpsilonGreedyAgent(
time_step_spec=environment.time_step_spec(),
action_spec=environment.action_spec(),
reward_network=network,
optimizer=tf.compat.v1.train.AdamOptimizer(learning_rate=LR),
epsilon=EPSILON,
accepts_per_arm_features=True,
emit_policy_info=policy_utilities.InfoFields.PREDICTED_REWARDS_MEAN)
optimal_reward_fn = functools.partial(optimal_reward,
hidden_param=HIDDEN_PARAM)
optimal_action_fn = functools.partial(optimal_action,
hidden_param=HIDDEN_PARAM)
regret_metric = tf_bandit_metrics.RegretMetric(optimal_reward_fn)
suboptimal_arms_metric = tf_bandit_metrics.SuboptimalArmsMetric(
optimal_action_fn)
trainer.train(root_dir=FLAGS.root_dir,
agent=agent,
environment=environment,
training_loops=TRAINING_LOOPS,
steps_per_loop=STEPS_PER_LOOP,
additional_metrics=[regret_metric, suboptimal_arms_metric])
def main(unused_argv):
tf.compat.v1.enable_v2_behavior() # The trainer only runs with V2 enabled.
with tf.device('/CPU:0'): # due to b/128333994
covertype_dataset = dataset_utilities.convert_covertype_dataset(
FLAGS.covertype_csv)
covertype_reward_distribution = tfd.Independent(
tfd.Deterministic(tf.eye(7)), reinterpreted_batch_ndims=2)
environment = ce.ClassificationBanditEnvironment(
covertype_dataset, covertype_reward_distribution, BATCH_SIZE)
optimal_reward_fn = functools.partial(
env_util.compute_optimal_reward_with_classification_environment,
environment=environment)
optimal_action_fn = functools.partial(
env_util.compute_optimal_action_with_classification_environment,
environment=environment)
if FLAGS.agent == 'LinUCB':
agent = lin_ucb_agent.LinearUCBAgent(
time_step_spec=environment.time_step_spec(),
action_spec=environment.action_spec(),
alpha=AGENT_ALPHA,
emit_log_probability=False,
dtype=tf.float32)
elif FLAGS.agent == 'LinTS':
agent = lin_ts_agent.LinearThompsonSamplingAgent(
time_step_spec=environment.time_step_spec(),
action_spec=environment.action_spec(),
alpha=AGENT_ALPHA,
dtype=tf.float32)
elif FLAGS.agent == 'epsGreedy':
network = q_network.QNetwork(
input_tensor_spec=environment.time_step_spec().observation,
action_spec=environment.action_spec(),
fc_layer_params=LAYERS)
agent = eps_greedy_agent.NeuralEpsilonGreedyAgent(
time_step_spec=environment.time_step_spec(),
action_spec=environment.action_spec(),
reward_network=network,
optimizer=tf.compat.v1.train.AdamOptimizer(learning_rate=LR),
epsilon=EPSILON)
regret_metric = tf_bandit_metrics.RegretMetric(optimal_reward_fn)
suboptimal_arms_metric = tf_bandit_metrics.SuboptimalArmsMetric(
optimal_action_fn)
trainer.train(
root_dir=FLAGS.root_dir,
agent=agent,
environment=environment,
training_loops=TRAINING_LOOPS,
steps_per_loop=STEPS_PER_LOOP,
additional_metrics=[regret_metric, suboptimal_arms_metric])
def main(unused_argv):
tf.enable_resource_variables()
with tf.device('/CPU:0'): # due to b/128333994
env = wheel_py_environment.WheelPyEnvironment(DELTA, MU_BASE, STD_BASE,
MU_HIGH, STD_HIGH,
BATCH_SIZE)
environment = tf_py_environment.TFPyEnvironment(env)
optimal_reward_fn = functools.partial(
environment_utilities.tf_wheel_bandit_compute_optimal_reward,
delta=DELTA,
mu_inside=MU_BASE[0],
mu_high=MU_HIGH)
optimal_action_fn = functools.partial(
environment_utilities.tf_wheel_bandit_compute_optimal_action,
delta=DELTA)
if FLAGS.agent == 'LinUCB':
agent = lin_ucb_agent.LinearUCBAgent(
time_step_spec=environment.time_step_spec(),
action_spec=environment.action_spec(),
alpha=AGENT_ALPHA,
dtype=tf.float32)
elif FLAGS.agent == 'LinTS':
agent = lin_ts_agent.LinearThompsonSamplingAgent(
time_step_spec=environment.time_step_spec(),
action_spec=environment.action_spec(),
alpha=AGENT_ALPHA,
dtype=tf.float32)
elif FLAGS.agent == 'epsGreedy':
network = q_network.QNetwork(
input_tensor_spec=environment.time_step_spec().observation,
action_spec=environment.action_spec(),
fc_layer_params=LAYERS)
agent = eps_greedy_agent.NeuralEpsilonGreedyAgent(
time_step_spec=environment.time_step_spec(),
action_spec=environment.action_spec(),
reward_network=network,
optimizer=tf.compat.v1.train.AdamOptimizer(learning_rate=LR),
epsilon=EPSILON)
regret_metric = tf_bandit_metrics.RegretMetric(optimal_reward_fn)
suboptimal_arms_metric = tf_bandit_metrics.SuboptimalArmsMetric(
optimal_action_fn)
trainer.train(
root_dir=FLAGS.root_dir,
agent=agent,
environment=environment,
training_loops=TRAINING_LOOPS,
steps_per_loop=STEPS_PER_LOOP,
additional_metrics=[regret_metric, suboptimal_arms_metric])
def main(unused_argv):
tf.compat.v1.enable_v2_behavior() # The trainer only runs with V2 enabled.
with tf.device('/CPU:0'): # due to b/128333994
mushroom_reward_distribution = (
dataset_utilities.mushroom_reward_distribution(
r_noeat=0.0,
r_eat_safe=5.0,
r_eat_poison_bad=-35.0,
r_eat_poison_good=5.0,
prob_poison_bad=0.5))
mushroom_dataset = (
dataset_utilities.convert_mushroom_csv_to_tf_dataset(
FLAGS.mushroom_csv))
environment = ce.ClassificationBanditEnvironment(
mushroom_dataset, mushroom_reward_distribution, BATCH_SIZE)
optimal_reward_fn = functools.partial(
env_util.compute_optimal_reward_with_classification_environment,
environment=environment)
optimal_action_fn = functools.partial(
env_util.compute_optimal_action_with_classification_environment,
environment=environment)
if FLAGS.agent == 'LinUCB':
agent = lin_ucb_agent.LinearUCBAgent(
time_step_spec=environment.time_step_spec(),
action_spec=environment.action_spec(),
alpha=AGENT_ALPHA,
gamma=0.95,
emit_log_probability=False,
dtype=tf.float32)
elif FLAGS.agent == 'LinTS':
agent = lin_ts_agent.LinearThompsonSamplingAgent(
time_step_spec=environment.time_step_spec(),
action_spec=environment.action_spec(),
alpha=AGENT_ALPHA,
gamma=0.95,
dtype=tf.float32)
regret_metric = tf_bandit_metrics.RegretMetric(optimal_reward_fn)
suboptimal_arms_metric = tf_bandit_metrics.SuboptimalArmsMetric(
optimal_action_fn)
trainer.train(
root_dir=FLAGS.root_dir,
agent=agent,
environment=environment,
training_loops=TRAINING_LOOPS,
steps_per_loop=STEPS_PER_LOOP,
additional_metrics=[regret_metric, suboptimal_arms_metric])
def main(unused_argv):
tf.compat.v1.enable_v2_behavior() # The trainer only runs with V2 enabled.
with tf.device('/CPU:0'): # due to b/128333994
observation_shape = [CONTEXT_DIM]
overall_shape = [BATCH_SIZE] + observation_shape
observation_distribution = tfd.Normal(loc=tf.zeros(overall_shape),
scale=tf.ones(overall_shape))
action_shape = [NUM_ACTIONS]
observation_to_reward_shape = observation_shape + action_shape
observation_to_reward_distribution = tfd.Normal(
loc=tf.zeros(observation_to_reward_shape),
scale=tf.ones(observation_to_reward_shape))
drift_distribution = tfd.Normal(loc=DRIFT_MEAN, scale=DRIFT_VARIANCE)
additive_reward_distribution = tfd.Normal(
loc=tf.zeros(action_shape),
scale=(REWARD_NOISE_VARIANCE * tf.ones(action_shape)))
environment_dynamics = dle.DriftingLinearDynamics(
observation_distribution, observation_to_reward_distribution,
drift_distribution, additive_reward_distribution)
environment = nse.NonStationaryStochasticEnvironment(
environment_dynamics)
if FLAGS.agent == 'LinUCB':
agent = lin_ucb_agent.LinearUCBAgent(
time_step_spec=environment.time_step_spec(),
action_spec=environment.action_spec(),
alpha=AGENT_ALPHA,
gamma=0.95,
emit_log_probability=False,
dtype=tf.float32)
elif FLAGS.agent == 'LinTS':
agent = lin_ts_agent.LinearThompsonSamplingAgent(
time_step_spec=environment.time_step_spec(),
action_spec=environment.action_spec(),
alpha=AGENT_ALPHA,
gamma=0.95,
dtype=tf.float32)
regret_metric = tf_bandit_metrics.RegretMetric(
environment.environment_dynamics.compute_optimal_reward)
suboptimal_arms_metric = tf_bandit_metrics.SuboptimalArmsMetric(
environment.environment_dynamics.compute_optimal_action)
trainer.train(
root_dir=FLAGS.root_dir,
agent=agent,
environment=environment,
training_loops=TRAINING_LOOPS,
steps_per_loop=STEPS_PER_LOOP,
additional_metrics=[regret_metric, suboptimal_arms_metric])
def main(unused_argv):
tf.compat.v1.enable_v2_behavior() # The trainer only runs with V2 enabled.
with tf.device('/CPU:0'): # due to b/128333994
action_reward_fns = (
environment_utilities.sliding_linear_reward_fn_generator(
CONTEXT_DIM, NUM_ACTIONS, REWARD_NOISE_VARIANCE))
env = sspe.StationaryStochasticPyEnvironment(functools.partial(
environment_utilities.context_sampling_fn,
batch_size=BATCH_SIZE,
context_dim=CONTEXT_DIM),
action_reward_fns,
batch_size=BATCH_SIZE)
environment = tf_py_environment.TFPyEnvironment(env)
optimal_reward_fn = functools.partial(
environment_utilities.tf_compute_optimal_reward,
per_action_reward_fns=action_reward_fns)
optimal_action_fn = functools.partial(
environment_utilities.tf_compute_optimal_action,
per_action_reward_fns=action_reward_fns)
q_net = q_network.QNetwork(environment.observation_spec(),
environment.action_spec(),
fc_layer_params=(50, 50))
agent = dqn_agent.DqnAgent(
environment.time_step_spec(),
environment.action_spec(),
q_network=q_net,
epsilon_greedy=0.1,
target_update_tau=0.05,
target_update_period=5,
optimizer=tf.compat.v1.train.AdamOptimizer(learning_rate=1e-2),
td_errors_loss_fn=common.element_wise_squared_loss)
regret_metric = tf_bandit_metrics.RegretMetric(optimal_reward_fn)
suboptimal_arms_metric = tf_bandit_metrics.SuboptimalArmsMetric(
optimal_action_fn)
trainer.train(
root_dir=FLAGS.root_dir,
agent=agent,
environment=environment,
training_loops=TRAINING_LOOPS,
steps_per_loop=STEPS_PER_LOOP,
additional_metrics=[regret_metric, suboptimal_arms_metric])
def main(unused_argv):
tf.compat.v1.enable_resource_variables()
with tf.device('/CPU:0'): # due to b/128333994
action_reward_fns = (
environment_utilities.sliding_linear_reward_fn_generator(
CONTEXT_DIM, NUM_ACTIONS, REWARD_NOISE_VARIANCE))
env = sspe.StationaryStochasticPyEnvironment(functools.partial(
environment_utilities.context_sampling_fn,
batch_size=BATCH_SIZE,
context_dim=CONTEXT_DIM),
action_reward_fns,
batch_size=BATCH_SIZE)
environment = tf_py_environment.TFPyEnvironment(env)
optimal_reward_fn = functools.partial(
environment_utilities.tf_compute_optimal_reward,
per_action_reward_fns=action_reward_fns)
optimal_action_fn = functools.partial(
environment_utilities.tf_compute_optimal_action,
per_action_reward_fns=action_reward_fns)
if FLAGS.agent == 'LinUCB':
agent = lin_ucb_agent.LinearUCBAgent(
time_step_spec=environment.time_step_spec(),
action_spec=environment.action_spec(),
alpha=AGENT_ALPHA,
dtype=tf.float32)
elif FLAGS.agent == 'LinTS':
agent = lin_ts_agent.LinearThompsonSamplingAgent(
time_step_spec=environment.time_step_spec(),
action_spec=environment.action_spec(),
alpha=AGENT_ALPHA,
dtype=tf.float32)
regret_metric = tf_bandit_metrics.RegretMetric(optimal_reward_fn)
suboptimal_arms_metric = tf_bandit_metrics.SuboptimalArmsMetric(
optimal_action_fn)
trainer.train(
root_dir=FLAGS.root_dir,
agent=agent,
environment=environment,
training_loops=TRAINING_LOOPS,
steps_per_loop=STEPS_PER_LOOP,
additional_metrics=[regret_metric, suboptimal_arms_metric])
def testAgentAndEnvironmentRuns(self, environment_name, agent_name):
batch_size = 8
training_loops = 3
steps_per_loop = 2
(environment, optimal_reward_fn,
optimal_action_fn) = get_environment_and_optimal_functions_by_name(
environment_name, batch_size)
agent = get_agent_by_name(agent_name, environment.time_step_spec(),
environment.action_spec())
regret_metric = tf_bandit_metrics.RegretMetric(optimal_reward_fn)
suboptimal_arms_metric = tf_bandit_metrics.SuboptimalArmsMetric(
optimal_action_fn)
trainer.train(
root_dir=tempfile.mkdtemp(dir=os.getenv('TEST_TMPDIR')),
agent=agent,
environment=environment,
training_loops=training_loops,
steps_per_loop=steps_per_loop,
additional_metrics=[regret_metric, suboptimal_arms_metric])
def main(unused_argv):
tf.compat.v1.enable_v2_behavior() # The trainer only runs with V2 enabled.
data_path = FLAGS.data_path
if not data_path:
raise ValueError('Please specify the location of the data file.')
if FLAGS.per_arm:
env = movielens_per_arm_py_environment.MovieLensPerArmPyEnvironment(
data_path,
RANK_K,
BATCH_SIZE,
num_actions=NUM_ACTIONS,
csv_delimiter='\t')
else:
env = movielens_py_environment.MovieLensPyEnvironment(
data_path,
RANK_K,
BATCH_SIZE,
num_movies=NUM_ACTIONS,
csv_delimiter='\t')
environment = tf_py_environment.TFPyEnvironment(env)
optimal_reward_fn = functools.partial(
environment_utilities.compute_optimal_reward_with_movielens_environment,
environment=environment)
optimal_action_fn = functools.partial(
environment_utilities.compute_optimal_action_with_movielens_environment,
environment=environment)
if FLAGS.agent == 'LinUCB':
agent = lin_ucb_agent.LinearUCBAgent(
time_step_spec=environment.time_step_spec(),
action_spec=environment.action_spec(),
tikhonov_weight=0.001,
alpha=AGENT_ALPHA,
dtype=tf.float32,
accepts_per_arm_features=FLAGS.per_arm)
elif FLAGS.agent == 'LinTS':
agent = lin_ts_agent.LinearThompsonSamplingAgent(
time_step_spec=environment.time_step_spec(),
action_spec=environment.action_spec(),
dtype=tf.float32,
accepts_per_arm_features=FLAGS.per_arm)
elif FLAGS.agent == 'epsGreedy':
if FLAGS.per_arm:
network = (
global_and_arm_feature_network
.create_feed_forward_dot_product_network(
environment.time_step_spec().observation,
global_layers=LAYERS,
arm_layers=LAYERS))
else:
network = q_network.QNetwork(
input_tensor_spec=environment.time_step_spec().observation,
action_spec=environment.action_spec(),
fc_layer_params=LAYERS)
agent = eps_greedy_agent.NeuralEpsilonGreedyAgent(
time_step_spec=environment.time_step_spec(),
action_spec=environment.action_spec(),
reward_network=network,
optimizer=tf.compat.v1.train.AdamOptimizer(learning_rate=LR),
epsilon=EPSILON,
emit_policy_info='predicted_rewards_mean',
info_fields_to_inherit_from_greedy=['predicted_rewards_mean'])
elif FLAGS.agent == 'DropoutTS':
train_step_counter = tf.compat.v1.train.get_or_create_global_step()
def dropout_fn():
return tf.math.maximum(
tf.math.reciprocal_no_nan(1.01 +
tf.cast(train_step_counter, tf.float32)),
0.0003)
agent = dropout_ts_agent.DropoutThompsonSamplingAgent(
time_step_spec=environment.time_step_spec(),
action_spec=environment.action_spec(),
dropout_rate=dropout_fn,
network_layers=LAYERS,
optimizer=tf.compat.v1.train.AdamOptimizer(learning_rate=LR))
regret_metric = tf_bandit_metrics.RegretMetric(optimal_reward_fn)
suboptimal_arms_metric = tf_bandit_metrics.SuboptimalArmsMetric(
optimal_action_fn)
trainer.train(
root_dir=FLAGS.root_dir,
agent=agent,
environment=environment,
training_loops=TRAINING_LOOPS,
steps_per_loop=STEPS_PER_LOOP,
additional_metrics=[regret_metric, suboptimal_arms_metric])
def main(unused_argv):
tf.compat.v1.enable_v2_behavior() # The trainer only runs with V2 enabled.
with tf.device('/CPU:0'): # due to b/128333994
env = wheel_py_environment.WheelPyEnvironment(DELTA, MU_BASE, STD_BASE,
MU_HIGH, STD_HIGH,
BATCH_SIZE)
environment = tf_py_environment.TFPyEnvironment(env)
optimal_reward_fn = functools.partial(
environment_utilities.tf_wheel_bandit_compute_optimal_reward,
delta=DELTA,
mu_inside=MU_BASE[0],
mu_high=MU_HIGH)
optimal_action_fn = functools.partial(
environment_utilities.tf_wheel_bandit_compute_optimal_action,
delta=DELTA)
network = q_network.QNetwork(
input_tensor_spec=environment.time_step_spec().observation,
action_spec=environment.action_spec(),
fc_layer_params=(LAYERS))
if FLAGS.agent == 'LinUCB':
agent = lin_ucb_agent.LinearUCBAgent(
time_step_spec=environment.time_step_spec(),
action_spec=environment.action_spec(),
alpha=AGENT_ALPHA,
dtype=tf.float32)
elif FLAGS.agent == 'LinTS':
agent = lin_ts_agent.LinearThompsonSamplingAgent(
time_step_spec=environment.time_step_spec(),
action_spec=environment.action_spec(),
alpha=AGENT_ALPHA,
dtype=tf.float32)
elif FLAGS.agent == 'epsGreedy':
agent = eps_greedy_agent.NeuralEpsilonGreedyAgent(
time_step_spec=environment.time_step_spec(),
action_spec=environment.action_spec(),
reward_network=network,
optimizer=tf.compat.v1.train.AdamOptimizer(learning_rate=LR),
epsilon=EPSILON)
elif FLAGS.agent == 'random':
agent = eps_greedy_agent.NeuralEpsilonGreedyAgent(
time_step_spec=environment.time_step_spec(),
action_spec=environment.action_spec(),
reward_network=network,
optimizer=tf.compat.v1.train.AdamOptimizer(learning_rate=LR),
epsilon=1.)
elif FLAGS.agent == 'Mix':
emit_policy_info = (
policy_utilities.InfoFields.PREDICTED_REWARDS_MEAN, )
agent_epsgreedy = eps_greedy_agent.NeuralEpsilonGreedyAgent(
time_step_spec=environment.time_step_spec(),
action_spec=environment.action_spec(),
reward_network=network,
optimizer=tf.compat.v1.train.AdamOptimizer(learning_rate=LR),
emit_policy_info=emit_policy_info,
epsilon=EPSILON)
agent_linucb = lin_ucb_agent.LinearUCBAgent(
time_step_spec=environment.time_step_spec(),
action_spec=environment.action_spec(),
alpha=AGENT_ALPHA,
emit_policy_info=emit_policy_info,
dtype=tf.float32)
agent_random = eps_greedy_agent.NeuralEpsilonGreedyAgent(
time_step_spec=environment.time_step_spec(),
action_spec=environment.action_spec(),
reward_network=network,
optimizer=tf.compat.v1.train.AdamOptimizer(learning_rate=LR),
emit_policy_info=emit_policy_info,
epsilon=1.)
agent_halfrandom = eps_greedy_agent.NeuralEpsilonGreedyAgent(
time_step_spec=environment.time_step_spec(),
action_spec=environment.action_spec(),
reward_network=network,
optimizer=tf.compat.v1.train.AdamOptimizer(learning_rate=LR),
emit_policy_info=emit_policy_info,
epsilon=0.5)
agent = exp3_mixture_agent.Exp3MixtureAgent(
(agent_epsgreedy, agent_linucb, agent_random,
agent_halfrandom))
regret_metric = tf_bandit_metrics.RegretMetric(optimal_reward_fn)
suboptimal_arms_metric = tf_bandit_metrics.SuboptimalArmsMetric(
optimal_action_fn)
trainer.train(
root_dir=FLAGS.root_dir,
agent=agent,
environment=environment,
training_loops=TRAINING_LOOPS,
steps_per_loop=STEPS_PER_LOOP,
additional_metrics=[regret_metric, suboptimal_arms_metric])
def main(unused_argv):
tf.compat.v1.enable_v2_behavior() # The trainer only runs with V2 enabled.
with tf.device('/CPU:0'): # due to b/128333994
if FLAGS.normalize_reward_fns:
action_reward_fns = (
environment_utilities.normalized_sliding_linear_reward_fn_generator(
CONTEXT_DIM, NUM_ACTIONS, REWARD_NOISE_VARIANCE))
else:
action_reward_fns = (
environment_utilities.sliding_linear_reward_fn_generator(
CONTEXT_DIM, NUM_ACTIONS, REWARD_NOISE_VARIANCE))
env = sspe.StationaryStochasticPyEnvironment(
functools.partial(
environment_utilities.context_sampling_fn,
batch_size=BATCH_SIZE,
context_dim=CONTEXT_DIM),
action_reward_fns,
batch_size=BATCH_SIZE)
mask_split_fn = None
if FLAGS.num_disabled_actions > 0:
mask_split_fn = lambda x: (x[0], x[1])
env = wrappers.ExtraDisabledActionsWrapper(env,
FLAGS.num_disabled_actions)
environment = tf_py_environment.TFPyEnvironment(env)
optimal_reward_fn = functools.partial(
environment_utilities.tf_compute_optimal_reward,
per_action_reward_fns=action_reward_fns)
optimal_action_fn = functools.partial(
environment_utilities.tf_compute_optimal_action,
per_action_reward_fns=action_reward_fns)
network_input_spec = environment.time_step_spec().observation
if FLAGS.num_disabled_actions > 0:
def _apply_only_to_observation(fn):
def result_fn(obs):
return fn(obs[0])
return result_fn
optimal_action_fn = _apply_only_to_observation(optimal_action_fn)
optimal_reward_fn = _apply_only_to_observation(optimal_reward_fn)
network_input_spec = network_input_spec[0]
network = q_network.QNetwork(
input_tensor_spec=network_input_spec,
action_spec=environment.action_spec(),
fc_layer_params=LAYERS)
if FLAGS.agent == 'LinUCB':
agent = lin_ucb_agent.LinearUCBAgent(
time_step_spec=environment.time_step_spec(),
action_spec=environment.action_spec(),
alpha=AGENT_ALPHA,
dtype=tf.float32,
observation_and_action_constraint_splitter=mask_split_fn)
elif FLAGS.agent == 'LinTS':
agent = lin_ts_agent.LinearThompsonSamplingAgent(
time_step_spec=environment.time_step_spec(),
action_spec=environment.action_spec(),
alpha=AGENT_ALPHA,
dtype=tf.float32,
observation_and_action_constraint_splitter=mask_split_fn)
elif FLAGS.agent == 'epsGreedy':
agent = neural_epsilon_greedy_agent.NeuralEpsilonGreedyAgent(
time_step_spec=environment.time_step_spec(),
action_spec=environment.action_spec(),
reward_network=network,
optimizer=tf.compat.v1.train.AdamOptimizer(learning_rate=LR),
epsilon=EPSILON,
observation_and_action_constraint_splitter=mask_split_fn)
elif FLAGS.agent == 'Mix':
assert FLAGS.num_disabled_actions == 0, (
'Extra actions with mixture agent not supported.')
emit_policy_info = policy_utilities.InfoFields.PREDICTED_REWARDS_MEAN
agent_linucb = lin_ucb_agent.LinearUCBAgent(
time_step_spec=environment.time_step_spec(),
action_spec=environment.action_spec(),
emit_policy_info=emit_policy_info,
alpha=AGENT_ALPHA,
dtype=tf.float32)
agent_lints = lin_ts_agent.LinearThompsonSamplingAgent(
time_step_spec=environment.time_step_spec(),
action_spec=environment.action_spec(),
emit_policy_info=emit_policy_info,
alpha=AGENT_ALPHA,
dtype=tf.float32)
agent_epsgreedy = neural_epsilon_greedy_agent.NeuralEpsilonGreedyAgent(
time_step_spec=environment.time_step_spec(),
action_spec=environment.action_spec(),
reward_network=network,
optimizer=tf.compat.v1.train.AdamOptimizer(learning_rate=LR),
emit_policy_info=emit_policy_info,
epsilon=EPSILON)
agent = exp3_mixture_agent.Exp3MixtureAgent(
(agent_linucb, agent_lints, agent_epsgreedy))
regret_metric = tf_bandit_metrics.RegretMetric(optimal_reward_fn)
suboptimal_arms_metric = tf_bandit_metrics.SuboptimalArmsMetric(
optimal_action_fn)
trainer.train(
root_dir=FLAGS.root_dir,
agent=agent,
environment=environment,
training_loops=TRAINING_LOOPS,
steps_per_loop=STEPS_PER_LOOP,
additional_metrics=[regret_metric, suboptimal_arms_metric])
def main(unused_argv):
tf.compat.v1.enable_v2_behavior() # The trainer only runs with V2 enabled.
class LinearNormalReward(object):
def __init__(self, theta):
self.theta = theta
def __call__(self, x):
mu = np.dot(x, self.theta)
return np.random.normal(mu, 1)
def _global_context_sampling_fn():
return np.random.randint(-10, 10, [4]).astype(np.float32)
def _arm_context_sampling_fn():
return np.random.randint(-2, 3, [5]).astype(np.float32)
reward_fn = LinearNormalReward(HIDDEN_PARAM)
observation_and_action_constraint_splitter = None
num_actions_fn = None
variable_action_method = bandit_spec_utils.VariableActionMethod.FIXED
if FLAGS.add_num_actions_feature:
num_actions_fn = lambda: NUM_ACTIONS
variable_action_method = (
bandit_spec_utils.VariableActionMethod.NUM_ACTIONS_FEATURE)
env = sspe.StationaryStochasticPerArmPyEnvironment(
_global_context_sampling_fn,
_arm_context_sampling_fn,
NUM_ACTIONS,
reward_fn,
num_actions_fn,
batch_size=BATCH_SIZE,
variable_action_method=variable_action_method)
environment = tf_py_environment.TFPyEnvironment(env)
if FLAGS.agent == 'LinUCB':
agent = lin_ucb_agent.LinearUCBAgent(
time_step_spec=environment.time_step_spec(),
action_spec=environment.action_spec(),
alpha=AGENT_ALPHA,
accepts_per_arm_features=True,
dtype=tf.float32)
elif FLAGS.agent == 'LinTS':
agent = lin_ts_agent.LinearThompsonSamplingAgent(
time_step_spec=environment.time_step_spec(),
action_spec=environment.action_spec(),
alpha=AGENT_ALPHA,
observation_and_action_constraint_splitter=(
observation_and_action_constraint_splitter),
accepts_per_arm_features=True,
dtype=tf.float32)
elif FLAGS.agent == 'epsGreedy':
obs_spec = environment.observation_spec()
if FLAGS.network == 'commontower':
network = (global_and_arm_feature_network.
create_feed_forward_common_tower_network(
obs_spec, (40, 30), (30, 40), (40, 20)))
elif FLAGS.network == 'dotproduct':
network = (global_and_arm_feature_network.
create_feed_forward_dot_product_network(
obs_spec, (4, 3, 6), (3, 4, 6)))
agent = neural_epsilon_greedy_agent.NeuralEpsilonGreedyAgent(
time_step_spec=environment.time_step_spec(),
action_spec=environment.action_spec(),
reward_network=network,
optimizer=tf.compat.v1.train.AdamOptimizer(learning_rate=LR),
epsilon=EPSILON,
observation_and_action_constraint_splitter=(
observation_and_action_constraint_splitter),
accepts_per_arm_features=True,
emit_policy_info=policy_utilities.InfoFields.PREDICTED_REWARDS_MEAN
)
elif FLAGS.agent == 'NeuralLinUCB':
obs_spec = environment.observation_spec()
network = (global_and_arm_feature_network.
create_feed_forward_common_tower_network(
obs_spec, (40, 30), (30, 40), (40, 20), ENCODING_DIM))
agent = neural_linucb_agent.NeuralLinUCBAgent(
time_step_spec=environment.time_step_spec(),
action_spec=environment.action_spec(),
encoding_network=network,
encoding_network_num_train_steps=EPS_PHASE_STEPS,
encoding_dim=ENCODING_DIM,
optimizer=tf.compat.v1.train.AdamOptimizer(learning_rate=LR),
alpha=1.0,
gamma=1.0,
epsilon_greedy=EPSILON,
accepts_per_arm_features=True,
debug_summaries=True,
summarize_grads_and_vars=True,
emit_policy_info=policy_utilities.InfoFields.PREDICTED_REWARDS_MEAN
)
def _all_rewards(observation, hidden_param):
"""Outputs rewards for all actions, given an observation."""
hidden_param = tf.cast(hidden_param, dtype=tf.float32)
global_obs = observation[bandit_spec_utils.GLOBAL_FEATURE_KEY]
per_arm_obs = observation[bandit_spec_utils.PER_ARM_FEATURE_KEY]
num_actions = tf.shape(per_arm_obs)[1]
tiled_global = tf.tile(tf.expand_dims(global_obs, axis=1),
[1, num_actions, 1])
concatenated = tf.concat([tiled_global, per_arm_obs], axis=-1)
rewards = tf.linalg.matvec(concatenated, hidden_param)
return rewards
def optimal_reward(observation, hidden_param):
return tf.reduce_max(_all_rewards(observation, hidden_param), axis=1)
def optimal_action(observation, hidden_param):
return tf.argmax(_all_rewards(observation, hidden_param),
axis=1,
output_type=tf.int32)
optimal_reward_fn = functools.partial(optimal_reward,
hidden_param=HIDDEN_PARAM)
optimal_action_fn = functools.partial(optimal_action,
hidden_param=HIDDEN_PARAM)
regret_metric = tf_bandit_metrics.RegretMetric(optimal_reward_fn)
suboptimal_arms_metric = tf_bandit_metrics.SuboptimalArmsMetric(
optimal_action_fn)
if FLAGS.drop_arm_obs:
drop_arm_feature_fn = functools.partial(
bandit_spec_utils.drop_arm_observation)
else:
drop_arm_feature_fn = None
trainer.train(root_dir=FLAGS.root_dir,
agent=agent,
environment=environment,
training_loops=TRAINING_LOOPS,
steps_per_loop=STEPS_PER_LOOP,
additional_metrics=[regret_metric, suboptimal_arms_metric],
training_data_spec_transformation_fn=drop_arm_feature_fn)
def main(unused_argv):
tf.compat.v1.enable_v2_behavior() # The trainer only runs with V2 enabled.
with tf.device('/CPU:0'): # due to b/128333994
action_reward_fns = (
environment_utilities.structured_linear_reward_fn_generator(
CONTEXT_DIM, NUM_ACTIONS, REWARD_NOISE_VARIANCE))
env = sspe.StationaryStochasticPyEnvironment(functools.partial(
environment_utilities.context_sampling_fn,
batch_size=BATCH_SIZE,
context_dim=CONTEXT_DIM),
action_reward_fns,
batch_size=BATCH_SIZE)
environment = tf_py_environment.TFPyEnvironment(env)
optimal_reward_fn = functools.partial(
environment_utilities.tf_compute_optimal_reward,
per_action_reward_fns=action_reward_fns)
optimal_action_fn = functools.partial(
environment_utilities.tf_compute_optimal_action,
per_action_reward_fns=action_reward_fns)
if FLAGS.agent == 'LinUCB':
agent = lin_ucb_agent.LinearUCBAgent(
time_step_spec=environment.time_step_spec(),
action_spec=environment.action_spec(),
alpha=AGENT_ALPHA,
dtype=tf.float32)
elif FLAGS.agent == 'epsGreedy':
laplacian_matrix = utils.build_laplacian_over_ordinal_integer_actions(
environment.action_spec())
network = q_network.QNetwork(
input_tensor_spec=environment.time_step_spec().observation,
action_spec=environment.action_spec(),
fc_layer_params=REWARD_NETWORK_LAYER_PARAMS)
agent = eps_greedy_agent.NeuralEpsilonGreedyAgent(
time_step_spec=environment.time_step_spec(),
action_spec=environment.action_spec(),
reward_network=network,
optimizer=tf.compat.v1.train.AdamOptimizer(
learning_rate=NN_LEARNING_RATE),
epsilon=EPSILON,
laplacian_matrix=laplacian_matrix,
laplacian_smoothing_weight=0.01)
elif FLAGS.agent == 'LinTS':
agent = lin_ts_agent.LinearThompsonSamplingAgent(
time_step_spec=environment.time_step_spec(),
action_spec=environment.action_spec(),
alpha=AGENT_ALPHA,
dtype=tf.float32)
regret_metric = tf_bandit_metrics.RegretMetric(optimal_reward_fn)
suboptimal_arms_metric = tf_bandit_metrics.SuboptimalArmsMetric(
optimal_action_fn)
trainer.train(
root_dir=FLAGS.root_dir,
agent=agent,
environment=environment,
training_loops=TRAINING_LOOPS,
steps_per_loop=STEPS_PER_LOOP,
additional_metrics=[regret_metric, suboptimal_arms_metric])
def main(unused_argv):
tf.compat.v1.enable_v2_behavior() # The trainer only runs with V2 enabled.
with tf.device('/CPU:0'): # due to b/128333994
if FLAGS.normalize_reward_fns:
action_reward_fns = (environment_utilities.
normalized_sliding_linear_reward_fn_generator(
CONTEXT_DIM, NUM_ACTIONS,
REWARD_NOISE_VARIANCE))
else:
action_reward_fns = (
environment_utilities.sliding_linear_reward_fn_generator(
CONTEXT_DIM, NUM_ACTIONS, REWARD_NOISE_VARIANCE))
env = sspe.StationaryStochasticPyEnvironment(functools.partial(
environment_utilities.context_sampling_fn,
batch_size=BATCH_SIZE,
context_dim=CONTEXT_DIM),
action_reward_fns,
batch_size=BATCH_SIZE)
mask_split_fn = None
if FLAGS.num_disabled_actions > 0:
mask_split_fn = lambda x: (x[0], x[1])
env = wrappers.ExtraDisabledActionsWrapper(
env, FLAGS.num_disabled_actions)
environment = tf_py_environment.TFPyEnvironment(env)
optimal_reward_fn = functools.partial(
environment_utilities.tf_compute_optimal_reward,
per_action_reward_fns=action_reward_fns)
optimal_action_fn = functools.partial(
environment_utilities.tf_compute_optimal_action,
per_action_reward_fns=action_reward_fns)
network_input_spec = environment.time_step_spec().observation
if FLAGS.num_disabled_actions > 0:
def _apply_only_to_observation(fn):
def result_fn(obs):
return fn(obs[0])
return result_fn
optimal_action_fn = _apply_only_to_observation(optimal_action_fn)
optimal_reward_fn = _apply_only_to_observation(optimal_reward_fn)
network_input_spec = network_input_spec[0]
network = q_network.QNetwork(input_tensor_spec=network_input_spec,
action_spec=environment.action_spec(),
fc_layer_params=LAYERS)
if FLAGS.agent == 'LinUCB':
agent = lin_ucb_agent.LinearUCBAgent(
time_step_spec=environment.time_step_spec(),
action_spec=environment.action_spec(),
alpha=AGENT_ALPHA,
dtype=tf.float32,
observation_and_action_constraint_splitter=mask_split_fn)
elif FLAGS.agent == 'LinTS':
agent = lin_ts_agent.LinearThompsonSamplingAgent(
time_step_spec=environment.time_step_spec(),
action_spec=environment.action_spec(),
alpha=AGENT_ALPHA,
dtype=tf.float32,
observation_and_action_constraint_splitter=mask_split_fn)
elif FLAGS.agent == 'epsGreedy':
agent = neural_epsilon_greedy_agent.NeuralEpsilonGreedyAgent(
time_step_spec=environment.time_step_spec(),
action_spec=environment.action_spec(),
reward_network=network,
optimizer=tf.compat.v1.train.AdamOptimizer(learning_rate=LR),
epsilon=EPSILON,
observation_and_action_constraint_splitter=mask_split_fn)
elif FLAGS.agent == 'Boltzmann':
train_step_counter = tf.compat.v1.train.get_or_create_global_step()
boundaries = [500]
temp_values = [1000.0, TEMPERATURE]
temp_schedule = tf.keras.optimizers.schedules.PiecewiseConstantDecay(
boundaries, temp_values)
def _temperature_fn():
# Any variable used in the function needs to be saved in the policy.
# This is true by default for the `train_step_counter`.
return temp_schedule(train_step_counter)
agent = neural_boltzmann_agent.NeuralBoltzmannAgent(
time_step_spec=environment.time_step_spec(),
action_spec=environment.action_spec(),
reward_network=network,
temperature=_temperature_fn,
optimizer=tf.compat.v1.train.AdamOptimizer(learning_rate=LR),
observation_and_action_constraint_splitter=mask_split_fn,
train_step_counter=train_step_counter)
# This is needed, otherwise the PolicySaver complains.
agent.policy.step = train_step_counter
elif FLAGS.agent == 'BoltzmannGumbel':
num_samples_list = [
tf.compat.v2.Variable(0,
dtype=tf.int32,
name='num_samples_{}'.format(k))
for k in range(NUM_ACTIONS)
]
agent = neural_boltzmann_agent.NeuralBoltzmannAgent(
time_step_spec=environment.time_step_spec(),
action_spec=environment.action_spec(),
reward_network=network,
boltzmann_gumbel_exploration_constant=250.0,
optimizer=tf.compat.v1.train.AdamOptimizer(learning_rate=LR),
observation_and_action_constraint_splitter=mask_split_fn,
num_samples_list=num_samples_list)
elif FLAGS.agent == 'Mix':
assert FLAGS.num_disabled_actions == 0, (
'Extra actions with mixture agent not supported.')
emit_policy_info = policy_utilities.InfoFields.PREDICTED_REWARDS_MEAN
agent_linucb = lin_ucb_agent.LinearUCBAgent(
time_step_spec=environment.time_step_spec(),
action_spec=environment.action_spec(),
emit_policy_info=emit_policy_info,
alpha=AGENT_ALPHA,
dtype=tf.float32)
agent_lints = lin_ts_agent.LinearThompsonSamplingAgent(
time_step_spec=environment.time_step_spec(),
action_spec=environment.action_spec(),
emit_policy_info=emit_policy_info,
alpha=AGENT_ALPHA,
dtype=tf.float32)
agent_epsgreedy = neural_epsilon_greedy_agent.NeuralEpsilonGreedyAgent(
time_step_spec=environment.time_step_spec(),
action_spec=environment.action_spec(),
reward_network=network,
optimizer=tf.compat.v1.train.AdamOptimizer(learning_rate=LR),
emit_policy_info=emit_policy_info,
epsilon=EPSILON)
agent = exp3_mixture_agent.Exp3MixtureAgent(
(agent_linucb, agent_lints, agent_epsgreedy))
regret_metric = tf_bandit_metrics.RegretMetric(optimal_reward_fn)
suboptimal_arms_metric = tf_bandit_metrics.SuboptimalArmsMetric(
optimal_action_fn)
trainer.train(
root_dir=FLAGS.root_dir,
agent=agent,
environment=environment,
training_loops=TRAINING_LOOPS,
steps_per_loop=STEPS_PER_LOOP,
additional_metrics=[regret_metric, suboptimal_arms_metric])