def learn(env,
network,
seed=None,
lr=5e-4,
total_timesteps=100000,
buffer_size=100000,
exploration_fraction=0.1,
exploration_final_eps=0.1,
train_freq=1,
batch_size=64,
print_freq=1,
eval_freq=2500,
checkpoint_freq=10000,
checkpoint_path=None,
learning_starts=1000,
gamma=1.0,
target_network_update_freq=500,
prioritized_replay=False,
prioritized_replay_alpha=0.6,
prioritized_replay_beta0=0.4,
prioritized_replay_beta_iters=None,
prioritized_replay_eps=1e-6,
param_noise=False,
callback=None,
load_path=None,
csv_path="results.csv",
method_type="baseline",
**network_kwargs):
"""Train a deepr model.
Parameters
-------
env: gym.Env
environment to train on
network: string or a function
neural network to use as a q function approximator. If string, has to be one of the names of registered models in baselines.common.models
(mlp, cnn, conv_only). If a function, should take an observation tensor and return a latent variable tensor, which
will be mapped to the Q function heads (see build_q_func in baselines.deepr.models for details on that)
seed: int or None
prng seed. The runs with the same seed "should" give the same results. If None, no seeding is used.
lr: float
learning rate for adam optimizer
total_timesteps: int
number of env steps to optimizer for
buffer_size: int
size of the replay buffer
exploration_fraction: float
fraction of entire training period over which the exploration rate is annealed
exploration_final_eps: float
final value of random action probability
train_freq: int
update the model every `train_freq` steps.
batch_size: int
size of a batch sampled from replay buffer for training
print_freq: int
how often to print out training progress
set to None to disable printing
checkpoint_freq: int
how often to save the model. This is so that the best version is restored
at the end of the training. If you do not wish to restore the best version at
the end of the training set this variable to None.
learning_starts: int
how many steps of the model to collect transitions for before learning starts
gamma: float
discount factor
target_network_update_freq: int
update the target network every `target_network_update_freq` steps.
prioritized_replay: True
if True prioritized replay buffer will be used.
prioritized_replay_alpha: float
alpha parameter for prioritized replay buffer
prioritized_replay_beta0: float
initial value of beta for prioritized replay buffer
prioritized_replay_beta_iters: int
number of iterations over which beta will be annealed from initial value
to 1.0. If set to None equals to total_timesteps.
prioritized_replay_eps: float
epsilon to add to the TD errors when updating priorities.
param_noise: bool
whether or not to use parameter space noise (https://arxiv.org/abs/1706.01905)
callback: (locals, globals) -> None
function called at every steps with state of the algorithm.
If callback returns true training stops.
load_path: str
path to load the model from. (default: None)
**network_kwargs
additional keyword arguments to pass to the network builder.
Returns
-------
act: ActWrapper
Wrapper over act function. Adds ability to save it and load it.
See header of baselines/deepr/categorical.py for details on the act function.
"""
# Create all the functions necessary to train the model
sess = get_session()
set_global_seeds(seed)
#q_func = build_q_func(network, **network_kwargs)
q_func = build_q_func(mlp(num_layers=4, num_hidden=64), **network_kwargs)
#q_func = build_q_func(mlp(num_layers=2, num_hidden=64, activation=tf.nn.relu), **network_kwargs)
# capture the shape outside the closure so that the env object is not serialized
# by cloudpickle when serializing make_obs_ph
observation_space = env.observation_space
def make_obs_ph(name):
return ObservationInput(observation_space, name=name)
act, train, update_target, debug = build_train(
make_obs_ph=make_obs_ph,
q_func=q_func,
num_actions=env.action_space.n,
optimizer=tf.train.AdamOptimizer(learning_rate=lr),
gamma=gamma,
grad_norm_clipping=10,
param_noise=param_noise)
act_params = {
'make_obs_ph': make_obs_ph,
'q_func': q_func,
'num_actions': env.action_space.n,
}
act = ActWrapper(act, act_params)
# Create the replay buffer
if prioritized_replay:
replay_buffer = PrioritizedReplayBuffer(buffer_size,
alpha=prioritized_replay_alpha)
if prioritized_replay_beta_iters is None:
prioritized_replay_beta_iters = total_timesteps
beta_schedule = LinearSchedule(prioritized_replay_beta_iters,
initial_p=prioritized_replay_beta0,
final_p=1.0)
else:
replay_buffer = ReplayBuffer(buffer_size)
beta_schedule = None
# Create the schedule for exploration starting from 1.
exploration = LinearSchedule(
schedule_timesteps=int(exploration_fraction * total_timesteps),
#initial_p=1.0,
initial_p=exploration_final_eps,
final_p=exploration_final_eps)
# Initialize the parameters and copy them to the target network.
U.initialize()
update_target()
eval_rewards = [0.0]
saved_mean_reward = None
obs = env.reset()
reset = True
with tempfile.TemporaryDirectory() as td:
td = checkpoint_path or td
model_file = os.path.join(td, "model")
model_saved = False
if tf.train.latest_checkpoint(td) is not None:
load_variables(model_file)
logger.log('Loaded model from {}'.format(model_file))
model_saved = True
elif load_path is not None:
load_variables(load_path)
logger.log('Loaded model from {}'.format(load_path))
csvfile = open(csv_path, 'w', newline='')
fieldnames = ['STEPS', 'REWARD']
writer = csv.DictWriter(csvfile, fieldnames=fieldnames)
writer.writeheader()
for t in range(total_timesteps + 1):
if callback is not None:
if callback(locals(), globals()):
break
# Take action and update exploration to the newest value
kwargs = {}
if not param_noise:
#update_eps = exploration.value(t)
update_eps = exploration_final_eps
update_param_noise_threshold = 0.
else:
update_eps = 0.
# Compute the threshold such that the KL divergence between perturbed and non-perturbed
# policy is comparable to eps-greedy exploration with eps = exploration.value(t).
# See Appendix C.1 in Parameter Space Noise for Exploration, Plappert et al., 2017
# for detailed explanation.
update_param_noise_threshold = -np.log(1. - exploration.value(
t) + exploration.value(t) / float(env.action_space.n))
kwargs['reset'] = reset
kwargs[
'update_param_noise_threshold'] = update_param_noise_threshold
kwargs['update_param_noise_scale'] = True
action_mask = get_mask(env, method_type)
a = act(np.array(obs)[None],
unused_actions_neginf_mask=action_mask,
update_eps=update_eps,
**kwargs)[0]
env_action = a
reset = False
new_obs, rew, done, _ = env.step(env_action)
eval_rewards[-1] += rew
action_mask_p = get_mask(env, method_type)
# Shaping
if method_type == 'shaping':
## look-ahead shaping
ap = act(np.array(new_obs)[None],
unused_actions_neginf_mask=action_mask_p,
stochastic=False)[0]
f = action_mask_p[ap] - action_mask[a]
rew = rew + f
# Store transition in the replay buffer.
#replay_buffer.add(obs, a, rew, new_obs, float(done), action_mask_p)
if method_type != 'shaping':
replay_buffer.add(obs, a, rew, new_obs, float(done),
np.zeros(env.action_space.n))
else:
replay_buffer.add(obs, a, rew, new_obs, float(done),
action_mask_p)
obs = new_obs
if t % eval_freq == 0:
eval_rewards.append(0.0)
if t > learning_starts and t % train_freq == 0:
# Minimize the error in Bellman's equation on a batch sampled from replay buffer.
if prioritized_replay:
experience = replay_buffer.sample(
batch_size, beta=beta_schedule.value(t))
(obses_t, actions, rewards, obses_tp1, dones, weights,
batch_idxes) = experience
else:
obses_t, actions, rewards, obses_tp1, dones, masks_tp1 = replay_buffer.sample(
batch_size)
weights, batch_idxes = np.ones_like(rewards), None
td_errors = train(obses_t, actions, rewards, obses_tp1, dones,
weights, masks_tp1)
if prioritized_replay:
new_priorities = np.abs(td_errors) + prioritized_replay_eps
replay_buffer.update_priorities(batch_idxes,
new_priorities)
if t > learning_starts and t % target_network_update_freq == 0:
# Update target network periodically.
update_target()
mean_eval_reward = round(np.mean(eval_rewards[-1 - print_freq:-1]),
1)
num_evals = len(eval_rewards)
if t > 0 and t % eval_freq == 0 and print_freq is not None and t % (
print_freq * eval_freq) == 0:
#if done and print_freq is not None and len(eval_rewards) % print_freq == 0:
logger.record_tabular("steps", t)
logger.record_tabular("evals", num_evals)
logger.record_tabular("average reward in this eval",
mean_eval_reward / (eval_freq))
logger.record_tabular("total reward in this eval",
mean_eval_reward)
logger.dump_tabular()
writer.writerow({
"STEPS": t,
"REWARD": mean_eval_reward / (eval_freq)
})
csvfile.flush()
if (checkpoint_freq is not None and t > learning_starts
and num_evals > 100 and t % checkpoint_freq == 0):
if saved_mean_reward is None or mean_eval_reward > saved_mean_reward:
if print_freq is not None:
logger.log(
"Saving model due to mean reward increase: {} -> {}"
.format(saved_mean_reward, mean_eval_reward))
save_variables(model_file)
model_saved = True
saved_mean_reward = mean_eval_reward
if model_saved:
if print_freq is not None:
logger.log("Restored model with mean reward: {}".format(
saved_mean_reward))
load_variables(model_file)
return act
def learn(env,
network,
seed=None,
lr=5e-5,
total_timesteps=100000,
buffer_size=500000,
exploration_fraction=0.1,
exploration_final_eps=0.01,
train_freq=1,
batch_size=32,
print_freq=10,
checkpoint_freq=100000,
checkpoint_path=None,
learning_starts=0,
gamma=0.99,
target_network_update_freq=10000,
prioritized_replay=True,
prioritized_replay_alpha=0.4,
prioritized_replay_beta0=0.6,
prioritized_replay_beta_iters=None,
prioritized_replay_eps=1e-3,
param_noise=False,
callback=None,
load_path=None,
load_idx=None,
demo_path=None,
n_step=10,
demo_prioritized_replay_eps=1.0,
pre_train_timesteps=750000,
epsilon_schedule="constant",
**network_kwargs):
# Create all the functions necessary to train the model
set_global_seeds(seed)
q_func = build_q_func(network, **network_kwargs)
with tf.device('/GPU:0'):
model = DQfD(q_func=q_func,
observation_shape=env.observation_space.shape,
num_actions=env.action_space.n,
lr=lr,
grad_norm_clipping=10,
gamma=gamma,
param_noise=param_noise)
# Load model from checkpoint
if load_path is not None:
load_path = osp.expanduser(load_path)
ckpt = tf.train.Checkpoint(model=model)
manager = tf.train.CheckpointManager(ckpt, load_path, max_to_keep=None)
if load_idx is None:
ckpt.restore(manager.latest_checkpoint)
print("Restoring from {}".format(manager.latest_checkpoint))
else:
ckpt.restore(manager.checkpoints[load_idx])
print("Restoring from {}".format(manager.checkpoints[load_idx]))
# Setup demo trajectory
assert demo_path is not None
with open(demo_path, "rb") as f:
trajectories = pickle.load(f)
# Create the replay buffer
replay_buffer = PrioritizedReplayBuffer(buffer_size,
prioritized_replay_alpha)
if prioritized_replay_beta_iters is None:
prioritized_replay_beta_iters = total_timesteps
beta_schedule = LinearSchedule(prioritized_replay_beta_iters,
initial_p=prioritized_replay_beta0,
final_p=1.0)
temp_buffer = deque(maxlen=n_step)
is_demo = True
for epi in trajectories:
for obs, action, rew, new_obs, done in epi:
obs, new_obs = np.expand_dims(
np.array(obs), axis=0), np.expand_dims(np.array(new_obs),
axis=0)
if n_step:
temp_buffer.append((obs, action, rew, new_obs, done, is_demo))
if len(temp_buffer) == n_step:
n_step_sample = get_n_step_sample(temp_buffer, gamma)
replay_buffer.demo_len += 1
replay_buffer.add(*n_step_sample)
else:
replay_buffer.demo_len += 1
replay_buffer.add(obs[0], action, rew, new_obs[0], float(done),
float(is_demo))
logger.log("trajectory length:", replay_buffer.demo_len)
# Create the schedule for exploration
if epsilon_schedule == "constant":
exploration = ConstantSchedule(exploration_final_eps)
else: # not used
exploration = LinearSchedule(schedule_timesteps=int(
exploration_fraction * total_timesteps),
initial_p=1.0,
final_p=exploration_final_eps)
model.update_target()
# ============================================== pre-training ======================================================
start = time()
num_episodes = 0
temp_buffer = deque(maxlen=n_step)
for t in tqdm(range(pre_train_timesteps)):
# sample and train
experience = replay_buffer.sample(batch_size,
beta=prioritized_replay_beta0)
batch_idxes = experience[-1]
if experience[6] is None: # for n_step = 0
obses_t, actions, rewards, obses_tp1, dones, is_demos = tuple(
map(tf.constant, experience[:6]))
obses_tpn, rewards_n, dones_n = None, None, None
weights = tf.constant(experience[-2])
else:
obses_t, actions, rewards, obses_tp1, dones, is_demos, obses_tpn, rewards_n, dones_n, weights = tuple(
map(tf.constant, experience[:-1]))
td_errors, n_td_errors, loss_dq, loss_n, loss_E, loss_l2, weighted_error = model.train(
obses_t, actions, rewards, obses_tp1, dones, is_demos, weights,
obses_tpn, rewards_n, dones_n)
# Update priorities
new_priorities = np.abs(td_errors) + np.abs(
n_td_errors) + demo_prioritized_replay_eps
replay_buffer.update_priorities(batch_idxes, new_priorities)
# Update target network periodically
if t > 0 and t % target_network_update_freq == 0:
model.update_target()
# Logging
elapsed_time = timedelta(time() - start)
if print_freq is not None and t % 10000 == 0:
logger.record_tabular("steps", t)
logger.record_tabular("episodes", num_episodes)
logger.record_tabular("mean 100 episode reward", 0)
logger.record_tabular("max 100 episode reward", 0)
logger.record_tabular("min 100 episode reward", 0)
logger.record_tabular("demo sample rate", 1)
logger.record_tabular("epsilon", 0)
logger.record_tabular("loss_td", np.mean(loss_dq.numpy()))
logger.record_tabular("loss_n_td", np.mean(loss_n.numpy()))
logger.record_tabular("loss_margin", np.mean(loss_E.numpy()))
logger.record_tabular("loss_l2", np.mean(loss_l2.numpy()))
logger.record_tabular("losses_all", weighted_error.numpy())
logger.record_tabular("% time spent exploring",
int(100 * exploration.value(t)))
logger.record_tabular("pre_train", True)
logger.record_tabular("elapsed time", elapsed_time)
logger.dump_tabular()
# ============================================== exploring =========================================================
sample_counts = 0
demo_used_counts = 0
episode_rewards = deque(maxlen=100)
this_episode_reward = 0.
best_score = 0.
saved_mean_reward = None
is_demo = False
obs = env.reset()
# Always mimic the vectorized env
obs = np.expand_dims(np.array(obs), axis=0)
reset = True
for t in tqdm(range(total_timesteps)):
if callback is not None:
if callback(locals(), globals()):
break
kwargs = {}
if not param_noise:
update_eps = tf.constant(exploration.value(t))
update_param_noise_threshold = 0.
else: # not used
update_eps = tf.constant(0.)
update_param_noise_threshold = -np.log(1. - exploration.value(t) +
exploration.value(t) /
float(env.action_space.n))
kwargs['reset'] = reset
kwargs[
'update_param_noise_threshold'] = update_param_noise_threshold
kwargs['update_param_noise_scale'] = True
action, epsilon, _, _ = model.step(tf.constant(obs),
update_eps=update_eps,
**kwargs)
action = action[0].numpy()
reset = False
new_obs, rew, done, _ = env.step(action)
# Store transition in the replay buffer.
new_obs = np.expand_dims(np.array(new_obs), axis=0)
if n_step:
temp_buffer.append((obs, action, rew, new_obs, done, is_demo))
if len(temp_buffer) == n_step:
n_step_sample = get_n_step_sample(temp_buffer, gamma)
replay_buffer.add(*n_step_sample)
else:
replay_buffer.add(obs[0], action, rew, new_obs[0], float(done), 0.)
obs = new_obs
# invert log scaled score for logging
this_episode_reward += np.sign(rew) * (np.exp(np.sign(rew) * rew) - 1.)
if done:
num_episodes += 1
obs = env.reset()
obs = np.expand_dims(np.array(obs), axis=0)
episode_rewards.append(this_episode_reward)
reset = True
if this_episode_reward > best_score:
best_score = this_episode_reward
ckpt = tf.train.Checkpoint(model=model)
manager = tf.train.CheckpointManager(ckpt,
'./best_model',
max_to_keep=1)
manager.save(t)
logger.log("saved best model")
this_episode_reward = 0.0
if t % train_freq == 0:
experience = replay_buffer.sample(batch_size,
beta=beta_schedule.value(t))
batch_idxes = experience[-1]
if experience[6] is None: # for n_step = 0
obses_t, actions, rewards, obses_tp1, dones, is_demos = tuple(
map(tf.constant, experience[:6]))
obses_tpn, rewards_n, dones_n = None, None, None
weights = tf.constant(experience[-2])
else:
obses_t, actions, rewards, obses_tp1, dones, is_demos, obses_tpn, rewards_n, dones_n, weights = tuple(
map(tf.constant, experience[:-1]))
td_errors, n_td_errors, loss_dq, loss_n, loss_E, loss_l2, weighted_error = model.train(
obses_t, actions, rewards, obses_tp1, dones, is_demos, weights,
obses_tpn, rewards_n, dones_n)
new_priorities = np.abs(td_errors) + np.abs(
n_td_errors
) + demo_prioritized_replay_eps * is_demos + prioritized_replay_eps * (
1. - is_demos)
replay_buffer.update_priorities(batch_idxes, new_priorities)
# for logging
sample_counts += batch_size
demo_used_counts += np.sum(is_demos)
if t % target_network_update_freq == 0:
# Update target network periodically.
model.update_target()
if t % checkpoint_freq == 0:
save_path = checkpoint_path
ckpt = tf.train.Checkpoint(model=model)
manager = tf.train.CheckpointManager(ckpt,
save_path,
max_to_keep=10)
manager.save(t)
logger.log("saved checkpoint")
elapsed_time = timedelta(time() - start)
if done and num_episodes > 0 and num_episodes % print_freq == 0:
logger.record_tabular("steps", t)
logger.record_tabular("episodes", num_episodes)
logger.record_tabular("mean 100 episode reward",
np.mean(episode_rewards))
logger.record_tabular("max 100 episode reward",
np.max(episode_rewards))
logger.record_tabular("min 100 episode reward",
np.min(episode_rewards))
logger.record_tabular("demo sample rate",
demo_used_counts / sample_counts)
logger.record_tabular("epsilon", epsilon.numpy())
logger.record_tabular("loss_td", np.mean(loss_dq.numpy()))
logger.record_tabular("loss_n_td", np.mean(loss_n.numpy()))
logger.record_tabular("loss_margin", np.mean(loss_E.numpy()))
logger.record_tabular("loss_l2", np.mean(loss_l2.numpy()))
logger.record_tabular("losses_all", weighted_error.numpy())
logger.record_tabular("% time spent exploring",
int(100 * exploration.value(t)))
logger.record_tabular("pre_train", False)
logger.record_tabular("elapsed time", elapsed_time)
logger.dump_tabular()
return model
def learn(env,
network,
seed=None,
lr=5e-4,
total_timesteps=100000,
buffer_size=50000,
exploration_fraction=0.1,
exploration_final_eps=0.02,
train_freq=1,
batch_size=32,
print_freq=100,
checkpoint_freq=10000,
checkpoint_path=None,
learning_starts=1000,
gamma=1.0,
target_network_update_freq=500,
prioritized_replay=False,
prioritized_replay_alpha=0.6,
prioritized_replay_beta0=0.4,
prioritized_replay_beta_iters=None,
prioritized_replay_eps=1e-6,
param_noise=False,
callback=None,
load_path=None,
train_mode = True,
**network_kwargs
):
"""Train a deepq model.
Parameters
-------
env: gym.Env
environment to train on
network: string or a function
neural network to use as a q function approximator. If string, has to be one of the names of registered models in baselines.common.models
(mlp, cnn, conv_only). If a function, should take an observation tensor and return a latent variable tensor, which
will be mapped to the Q function heads (see build_q_func in baselines.deepq.models for details on that)
seed: int or None
prng seed. The runs with the same seed "should" give the same results. If None, no seeding is used.
lr: float
learning rate for adam optimizer
total_timesteps: int
number of env steps to optimizer for
buffer_size: int
size of the replay buffer
exploration_fraction: float
fraction of entire training period over which the exploration rate is annealed
exploration_final_eps: float
final value of random action probability
train_freq: int
update the model every `train_freq` steps.
batch_size: int
size of a batch sampled from replay buffer for training
print_freq: int
how often to print out training progress
set to None to disable printing
checkpoint_freq: int
how often to save the model. This is so that the best version is restored
at the end of the training. If you do not wish to restore the best version at
the end of the training set this variable to None.
learning_starts: int
how many steps of the model to collect transitions for before learning starts
gamma: float
discount factor
target_network_update_freq: int
update the target network every `target_network_update_freq` steps.
prioritized_replay: True
if True prioritized replay buffer will be used.
prioritized_replay_alpha: float
alpha parameter for prioritized replay buffer
prioritized_replay_beta0: float
initial value of beta for prioritized replay buffer
prioritized_replay_beta_iters: int
number of iterations over which beta will be annealed from initial value
to 1.0. If set to None equals to total_timesteps.
prioritized_replay_eps: float
epsilon to add to the TD errors when updating priorities.
param_noise: bool
whether or not to use parameter space noise (https://arxiv.org/abs/1706.01905)
callback: (locals, globals) -> None
function called at every steps with state of the algorithm.
If callback returns true training stops.
load_path: str
path to load the model from. (default: None)
**network_kwargs
additional keyword arguments to pass to the network builder.
Returns
-------
act: ActWrapper
Wrapper over act function. Adds ability to save it and load it.
See header of baselines/deepq/categorical.py for details on the act function.
"""
# Examine environment parameters
print(str(env))
# Set the default brain to work with
default_brain = env.brain_names[0]
brain = env.brains[default_brain]
num_actions=brain.vector_action_space_size[0]
# Create all the functions necessary to train the model
sess = get_session()
#set_global_seeds(seed)
q_func = build_q_func(network, **network_kwargs)
# capture the shape outside the closure so that the env object is not serialized
# by cloudpickle when serializing make_obs_ph
#observation_space = env.observation_space
env_info = env.reset(train_mode=train_mode)[default_brain]
state = get_obs_state_lidar(env_info)
observation_space=state.copy()
#def make_obs_ph(name,Num_action):
# tf.placeholder(shape=(None,) + state.shape, dtype=state.dtype, name='st')
# return tf.placeholder(tf.float32, shape = [None, Num_action],name=name)
def make_obs_ph(name):
return ObservationInput(observation_space, name=name)
act, train, update_target, debug =build_train(
make_obs_ph=make_obs_ph,
q_func=q_func,
num_actions=num_actions,
optimizer=tf.train.AdamOptimizer(learning_rate=lr),
gamma=gamma,
grad_norm_clipping=10,
param_noise=param_noise
)
act_params = {
'make_obs_ph': make_obs_ph,
'q_func': q_func,
'num_actions': num_actions,
}
act = ActWrapper(act, act_params)
# Create the replay buffer
if prioritized_replay:
replay_buffer = PrioritizedReplayBuffer(buffer_size, alpha=prioritized_replay_alpha)
if prioritized_replay_beta_iters is None:
prioritized_replay_beta_iters = total_timesteps
beta_schedule = LinearSchedule(prioritized_replay_beta_iters,
initial_p=prioritized_replay_beta0,
final_p=1.0)
else:
replay_buffer = ReplayBuffer(buffer_size)
beta_schedule = None
# Create the schedule for exploration starting from 1.
exploration = LinearSchedule(schedule_timesteps=int(exploration_fraction * total_timesteps),
initial_p=1.0,
final_p=exploration_final_eps)
# Initialize the parameters and copy them to the target network.
initialize()
update_target()
episode_rewards = [0.0]
saved_mean_reward = None
obs = env.reset()
reset = True
with tempfile.TemporaryDirectory() as td:
td = checkpoint_path or td
model_file = os.path.join(td, "model")
model_saved = False
if tf.train.latest_checkpoint(td) is not None:
load_variables(model_file)
logger.log('Loaded model from {}'.format(model_file))
model_saved = True
elif load_path is not None:
load_variables(load_path)
logger.log('Loaded model from {}'.format(load_path))
for t in range(total_timesteps):
if callback is not None:
if callback(locals(), globals()):
break
# Take action and update exploration to the newest value
kwargs = {}
if not param_noise:
update_eps = exploration.value(t)
update_param_noise_threshold = 0.
else:
update_eps = 0.
# Compute the threshold such that the KL divergence between perturbed and non-perturbed
# policy is comparable to eps-greedy exploration with eps = exploration.value(t).
# See Appendix C.1 in Parameter Space Noise for Exploration, Plappert et al., 2017
# for detailed explanation.
update_param_noise_threshold = -np.log(1. - exploration.value(t) + exploration.value(t) / float(num_actions))
kwargs['reset'] = reset
kwargs['update_param_noise_threshold'] = update_param_noise_threshold
kwargs['update_param_noise_scale'] = True
action = act(np.array(obs)[None], update_eps=update_eps, **kwargs)[0]
env_action = action
reset = False
new_obs, rew, done, _ = env.step(env_action)
# Store transition in the replay buffer.
replay_buffer.add(obs, action, rew, new_obs, float(done))
obs = new_obs
episode_rewards[-1] += rew
if done:
obs = env.reset()
episode_rewards.append(0.0)
reset = True
if t > learning_starts and t % train_freq == 0:
# Minimize the error in Bellman's equation on a batch sampled from replay buffer.
if prioritized_replay:
experience = replay_buffer.sample(batch_size, beta=beta_schedule.value(t))
(obses_t, actions, rewards, obses_tp1, dones, weights, batch_idxes) = experience
else:
obses_t, actions, rewards, obses_tp1, dones = replay_buffer.sample(batch_size)
weights, batch_idxes = np.ones_like(rewards), None
td_errors = train(obses_t, actions, rewards, obses_tp1, dones, weights)
if prioritized_replay:
new_priorities = np.abs(td_errors) + prioritized_replay_eps
replay_buffer.update_priorities(batch_idxes, new_priorities)
if t > learning_starts and t % target_network_update_freq == 0:
# Update target network periodically.
update_target()
mean_100ep_reward = round(np.mean(episode_rewards[-101:-1]), 1)
num_episodes = len(episode_rewards)
if done and print_freq is not None and len(episode_rewards) % print_freq == 0:
logger.record_tabular("steps", t)
logger.record_tabular("episodes", num_episodes)
logger.record_tabular("mean 100 episode reward", mean_100ep_reward)
logger.record_tabular("% time spent exploring", int(100 * exploration.value(t)))
logger.dump_tabular()
if (checkpoint_freq is not None and t > learning_starts and
num_episodes > 100 and t % checkpoint_freq == 0):
if saved_mean_reward is None or mean_100ep_reward > saved_mean_reward:
if print_freq is not None:
logger.log("Saving model due to mean reward increase: {} -> {}".format(
saved_mean_reward, mean_100ep_reward))
save_variables(model_file)
model_saved = True
saved_mean_reward = mean_100ep_reward
if model_saved:
if print_freq is not None:
logger.log("Restored model with mean reward: {}".format(saved_mean_reward))
load_variables(model_file)
return act