def build_train(make_obs_ph, model, num_actions, optimizer_f, grad_norm_clipping=None, gamma=1.0,
double_q=False, scope="deepq", reuse=None, param_noise=False, param_noise_filter_func=None, test_eps=0.05,
learning_rate = 0.001, learning_rate_decay_factor=0.99, learning_rate_growth_factor=1.001):
"""Creates the train function:
Parameters
----------
make_obs_ph: str -> tf.placeholder or TfInput
a function that takes a name and creates a placeholder of input with that name
q_func: (tf.Variable, int, str, bool) -> tf.Variable
the model that takes the following inputs:
observation_in: object
the output of observation placeholder
num_actions: int
number of actions
scope: str
reuse: bool
should be passed to outer variable scope
and returns a tensor of shape (batch_size, num_actions) with values of every action.
num_actions: int
number of actions
reuse: bool
whether or not to reuse the graph variables
optimizer: tf.train.Optimizer
optimizer to use for the Q-learning objective.
grad_norm_clipping: float or None
clip gradient norms to this value. If None no clipping is performed.
gamma: float
discount rate.
double_q: bool
if true will use Double Q Learning (https://arxiv.org/abs/1509.06461).
In general it is a good idea to keep it enabled.
scope: str or VariableScope
optional scope for variable_scope.
reuse: bool or None
whether or not the variables should be reused. To be able to reuse the scope must be given.
param_noise: bool
whether or not to use parameter space noise (https://arxiv.org/abs/1706.01905)
param_noise_filter_func: tf.Variable -> bool
function that decides whether or not a variable should be perturbed. Only applicable
if param_noise is True. If set to None, default_param_noise_filter is used by default.
Returns
-------
act: (tf.Variable, bool, float) -> tf.Variable
function to select and action given observation.
` See the top of the file for details.
train: (object, np.array, np.array, object, np.array, np.array) -> np.array
optimize the error in Bellman's equation.
` See the top of the file for details.
update_target: () -> ()
copy the parameters from optimized Q function to the target Q function.
` See the top of the file for details.
debug: {str: function}
a bunch of functions to print debug data like q_values.
"""
if param_noise:
act_f = build_act_with_param_noise(make_obs_ph, model, num_actions, scope=scope, reuse=reuse,
param_noise_filter_func=param_noise_filter_func)
else:
act_f = build_act(make_obs_ph, model, num_actions, scope=scope, reuse=reuse)
act_greedy = build_act_greedy(make_obs_ph, model, num_actions, scope=scope, reuse=True, eps=test_eps)
with tf.variable_scope(scope, reuse=reuse):
# set up placeholders
obs_t_input = make_obs_ph("obs_t")
act_t_ph = tf.placeholder(tf.int32, [None], name="action")
rew_t_ph = tf.placeholder(tf.float32, [None], name="reward")
obs_tp1_input = make_obs_ph("obs_tp1")
done_mask_ph = tf.placeholder(tf.float32, [None], name="done")
importance_weights_ph = tf.placeholder(tf.float32, [None], name="weight")
# Learning rate adjustment
lr = tf.Variable(float(learning_rate), trainable=False, dtype = tf.float32)
learning_rate_decay_op = lr.assign(tf.clip_by_value(lr*learning_rate_decay_factor, 1e-5, 1e-3))
learning_rate_growth_op = lr.assign(tf.clip_by_value(lr*learning_rate_growth_factor, 1e-5, 1e-3))
optimizer = optimizer_f(learning_rate = lr)
# q network evaluation
atom_t = model(obs_t_input.get(), num_outputs, scope="atom_func", reuse=True) # reuse parameters from act
atom_func_vars = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, scope=tf.get_variable_scope().name + "/atom_func")
atom_p_t = tf.nn.softmax(atom_t)
# target q network evalution
atom_tp1 = model(obs_tp1_input.get(), num_outputs, scope="target_atom_func")
target_atom_func_vars = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, scope=tf.get_variable_scope().name + "/target_atom_func")
atom_p_tp1 = tf.nn.softmax(atom_tp1)
m_vec = tf.constant(0.0, dtype=tf.float32, shape=(num_atoms))
for j in range(num_atoms):
Tz_j = tf.clip(rew_t_ph + gamma * (V_min + j * del_z), V_min, V_max)
b_j = (Tz_j - V_min)/del_z
l = tf.astype(tf.math.floor(b_j), tf.int32)
u = tf.astype(tf.math.ceil(b_j), tf.int32)
m_vec[l] = m_vec[l] +
cem_loss = tf.reduce_sum(tf.math.multiply(m, tf.log(atom_p)))
# q scores for actions which we know were selected in the given state.
q_t_selected = tf.reduce_sum(q_t * tf.one_hot(act_t_ph, num_actions), 1)
# compute estimate of best possible value starting from state at t + 1
q_tp1_best = tf.reduce_max(q_tp1, 1)
q_tp1_best_masked = (1.0 - done_mask_ph) * q_tp1_best
# compute RHS of bellman equation
q_t_selected_target = rew_t_ph + gamma * q_tp1_best_masked
# compute the error (potentially clipped)
td_error = q_t_selected - tf.stop_gradient(q_t_selected_target)
errors = U.huber_loss(td_error)
weighted_error = tf.reduce_mean(importance_weights_ph * errors)
# compute optimization op (potentially with gradient clipping)
if grad_norm_clipping is not None:
gradients = optimizer.compute_gradients(weighted_error, var_list=q_func_vars)
for i, (grad, var) in enumerate(gradients):
if grad is not None:
gradients[i] = (tf.clip_by_norm(grad, grad_norm_clipping), var)
optimize_expr = optimizer.apply_gradients(gradients)
else:
optimize_expr = optimizer.minimize(weighted_error, var_list=q_func_vars)
# update_target_fn will be called periodically to copy Q network to target Q network
update_target_expr = []
for var, var_target in zip(sorted(q_func_vars, key=lambda v: v.name),
sorted(target_q_func_vars, key=lambda v: v.name)):
update_target_expr.append(var_target.assign(var))
update_target_expr = tf.group(*update_target_expr)
# Create callable functions
train = U.function(
inputs=[
obs_t_input,
act_t_ph,
rew_t_ph,
obs_tp1_input,
done_mask_ph,
importance_weights_ph
],
outputs=[td_error, lr],
updates=[optimize_expr]
)
update_target = U.function([], [], updates=[update_target_expr])
q_values = U.function(inputs=[obs_t_input], outputs=q_t)
return act_f, act_greedy, q_values, train, update_target, learning_rate_decay_op, learning_rate_growth_op, {'q_values': q_values}
def build_train(make_obs_ph,
q_func,
num_actions,
optimizer_f,
grad_norm_clipping=None,
gamma=0.9,
scope="deepadfq",
reuse=None,
varTH=1e-05,
test_eps=0.05,
act_policy='egreedy',
learning_rate=0.001,
learning_rate_decay_factor=0.99,
learning_rate_growth_factor=1.001):
"""Creates the train function:
Parameters
----------
make_obs_ph: str -> tf.placeholder or TfInput
a function that takes a name and creates a placeholder of input with that name
q_func: (tf.Variable, int, str, bool) -> tf.Variable
the model that takes the following inputs:
observation_in: object
the output of observation placeholder
num_actions: int
number of actions
scope: str
reuse: bool
should be passed to outer variable scope
and returns a tensor of shape (batch_size, num_actions) with values of every action.
num_actions: int
number of actions
reuse: bool
whether or not to reuse the graph variables
optimizer: tf.train.Optimizer
optimizer to use for the Q-learning objective.
grad_norm_clipping: float or None
clip gradient norms to this value. If None no clipping is performed.
gamma: float
discount rate.
scope: str or VariableScope
optional scope for variable_scope.
reuse: bool or None
whether or not the variables should be reused. To be able to reuse the scope must be given.
param_noise: bool
whether or not to use parameter space noise (https://arxiv.org/abs/1706.01905)
param_noise_filter_func: tf.Variable -> bool
function that decides whether or not a variable should be perturbed. Only applicable
if param_noise is True. If set to None, default_param_noise_filter is used by default.
varTH : float
variance threshold
test_eps : float
epsilon value for epsilon-greedy method in evaluation
act_policy : str
either 'egreedy' or 'bayesian' for action policy
Returns
-------
act: (tf.Variable, bool, float) -> tf.Variable
function to select and action given observation.
` See the top of the file for details.
train: (object, np.array, np.array, object, np.array, np.array) -> np.array
optimize the error in Bellman's equation.
` See the top of the file for details.
update_target: () -> ()
copy the parameters from optimized Q function to the target Q function.
` See the top of the file for details.
debug: {str: function}
a bunch of functions to print debug data like q_values.
"""
if act_policy == 'egreedy':
act_f = build_act(make_obs_ph,
q_func,
num_actions,
scope=scope,
reuse=reuse)
elif act_policy == 'bayesian':
act_f = build_act_bayesian(make_obs_ph,
q_func,
num_actions,
scope=scope,
reuse=reuse)
else:
raise ValueError(
"Please choose either egreedy or bayesian for action policy.")
act_test = build_act_greedy(make_obs_ph,
q_func,
num_actions,
scope=scope,
reuse=True,
eps=test_eps)
#act_test = build_act_bayesian(make_obs_ph, q_func, num_actions, scope=scope, reuse=True)
sdTH = np.sqrt(varTH, dtype=np.float32)
with tf.variable_scope(scope, reuse=reuse):
# set up placeholders
obs_t_input = make_obs_ph("obs_t")
act_t_ph = tf.placeholder(tf.int32, [None], name="action")
rew_t_ph = tf.placeholder(tf.float32, [None], name="reward")
obs_tp1_input = make_obs_ph("obs_tp1")
done_mask_ph = tf.placeholder(tf.float32, [None], name="done")
importance_weights_ph = tf.placeholder(tf.float32, [None],
name="weight")
# Learning rate adjustment
lr = tf.Variable(float(learning_rate),
trainable=False,
dtype=tf.float32)
learning_rate_decay_op = lr.assign(
tf.clip_by_value(lr * learning_rate_decay_factor, 1e-5, 1e-3))
learning_rate_growth_op = lr.assign(
tf.clip_by_value(lr * learning_rate_growth_factor, 1e-5, 1e-3))
optimizer = optimizer_f(learning_rate=lr)
target_means = tf.placeholder(tf.float32, [None], name="target_means")
target_sd = tf.placeholder(tf.float32, [None], name="target_sd")
# q network evaluation
q_t = q_func(obs_t_input.get(),
num_actions * 2,
scope="q_func",
reuse=True) # reuse parameters from act
q_func_vars = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES,
scope=tf.get_variable_scope().name +
"/q_func")
# target q network evalution
q_tp1 = q_func(obs_tp1_input.get(),
num_actions * 2,
scope="target_q_func")
target_q_func_vars = tf.get_collection(
tf.GraphKeys.GLOBAL_VARIABLES,
scope=tf.get_variable_scope().name + "/target_q_func")
mean_values = q_t[:, :num_actions]
rho_values = q_t[:, num_actions:]
mean_selected = tf.reduce_sum(
mean_values * tf.one_hot(act_t_ph, num_actions, dtype=tf.float32),
1)
rho_selected = tf.reduce_sum(
rho_values * tf.one_hot(act_t_ph, num_actions, dtype=tf.float32),
1)
sd_selected = tf.exp(-rho_selected)
mean_error = mean_selected - tf.stop_gradient(target_means)
#sd_error = sd_selected - tf.stop_gradient(target_sd)
sd_error = tf.log(sd_selected) - tf.log(tf.stop_gradient(target_sd))
huber_loss = U.huber_loss(mean_error) + U.huber_loss(sd_error)
weighted_loss = tf.reduce_mean(huber_loss * importance_weights_ph)
#kl_loss = tf.contrib.distributions.kl_divergence(
# tf.distributions.Normal(loc=target_means, scale=target_sd),
# tf.distributions.Normal(loc=mean_selected, scale=sd_selected),
# name='kl_loss')
#weighted_loss = tf.reduce_mean(kl_loss * importance_weights_ph)
if grad_norm_clipping is not None:
gradients = optimizer.compute_gradients(weighted_loss,
var_list=q_func_vars)
for i, (grad, var) in enumerate(gradients):
if grad is not None:
gradients[i] = (tf.clip_by_norm(grad,
grad_norm_clipping), var)
optimize_expr = optimizer.apply_gradients(gradients)
else:
optimize_expr = optimizer.minimize(weighted_loss,
var_list=q_func_vars)
update_target_expr = []
for var, var_target in zip(
sorted(q_func_vars, key=lambda v: v.name),
sorted(target_q_func_vars, key=lambda v: v.name)):
update_target_expr.append(var_target.assign(var))
update_target_expr = tf.group(*update_target_expr)
# Create callable functions
train = U.function(
inputs=[
obs_t_input,
act_t_ph,
target_means,
target_sd,
importance_weights_ph,
],
outputs=[tf.reduce_mean(huber_loss), mean_error, sd_error, lr],
updates=[optimize_expr])
update_target = U.function([], [], updates=[update_target_expr])
q_target_vals = U.function(inputs=[obs_tp1_input], outputs=[q_tp1])
return act_f, act_test, q_target_vals, train, update_target, learning_rate_decay_op, learning_rate_growth_op
def build_train(make_obs_ph, q_func, num_actions, optimizer_f,
grad_norm_clipping=None, gamma=1.0, scope="setdeepq", reuse=None,
test_eps=0.05, lr_init = 0.001, lr_period_steps=50000, tau=0.05):
"""Creates the train function:
Parameters
----------
make_obs_ph: str -> tf.placeholder or TfInput
a function that takes a name and creates a placeholder of input with that name
q_func: (tf.Variable, int, str, bool) -> tf.Variable
the model that takes the following inputs:
observation_in: object
the output of observation placeholder
num_actions: int
number of actions
scope: str
reuse: bool
should be passed to outer variable scope
and returns a tensor of shape (batch_size, num_actions) with values of every action.
num_actions: int
number of actions
reuse: bool
whether or not to reuse the graph variables
optimizer: tf.train.Optimizer
optimizer to use for the Q-learning objective.
grad_norm_clipping: float or None
clip gradient norms to this value. If None no clipping is performed.
gamma: float
discount rate.
double_q: bool
if true will use Double Q Learning (https://arxiv.org/abs/1509.06461).
In general it is a good idea to keep it enabled.
scope: str or VariableScope
optional scope for variable_scope.
reuse: bool or None
whether or not the variables should be reused. To be able to reuse the scope must be given.
param_noise: bool
whether or not to use parameter space noise (https://arxiv.org/abs/1706.01905)
param_noise_filter_func: tf.Variable -> bool
function that decides whether or not a variable should be perturbed. Only applicable
if param_noise is True. If set to None, default_param_noise_filter is used by default.
lr_init : float
initial learning rate
lr_decay_factor : float
learning rate decay factor. It should be equal to or smaller than 1.0.
lr_growth_factor : float
learning rate growth factor. It should be equal to or larger than 1.0.
tau : float
parameter for the soft target network update. tau <= 1.0 and 1.0 for
the hard update.
Returns
-------
act: (tf.Variable, bool, float) -> tf.Variable
function to select and action given observation.
` See the top of the file for details.
train: (object, np.array, np.array, object, np.array, np.array) -> np.array
optimize the error in Bellman's equation.
` See the top of the file for details.
update_target: () -> ()
copy the parameters from optimized Q function to the target Q function.
` See the top of the file for details.
debug: {str: function}
a bunch of functions to print debug data like q_values.
"""
# Build action graphs
act_f = build_act(make_obs_ph, q_func, num_actions, scope=scope, reuse=reuse)
act_greedy = build_act_greedy(make_obs_ph, q_func, num_actions, scope=scope, reuse=True, eps=test_eps)
with tf.compat.v1.variable_scope(scope, reuse=reuse):
# set up placeholders
obs_t_input = make_obs_ph("obs_t")
act_t_ph = tf.compat.v1.placeholder(tf.int32, [None], name="action")
rew_t_ph = tf.compat.v1.placeholder(tf.float32, [None], name="reward")
obs_tp1_input = make_obs_ph("obs_tp1")
done_mask_ph = tf.compat.v1.placeholder(tf.float32, [None], name="done")
importance_weights_ph = tf.compat.v1.placeholder(tf.float32, [None], name="weight")
iteration = tf.compat.v1.placeholder(tf.float32, name="iteration")
# Cosine learning rate adjustment
lr = tf.Variable(float(lr_init), trainable=False, dtype = tf.float32, name='lr')
lr = tf.clip_by_value(0.0005*tf.math.cos(math.pi*iteration/lr_period_steps)+0.000501, 1e-6, 1e-3)
optimizer = optimizer_f(learning_rate = lr)
# q network evaluation
q1_t = q_func(obs_t_input.get(), num_actions, scope="q1_func", reuse=True) # reuse q1 parameters from act
q1_func_vars = tf.compat.v1.get_collection(tf.compat.v1.GraphKeys.TRAINABLE_VARIABLES, scope=tf.compat.v1.get_variable_scope().name + "/q1_func")
q2_t = q_func(obs_t_input.get(), num_actions, scope="q2_func", reuse=True) # reuse q2 parameters from act
q2_func_vars = tf.compat.v1.get_collection(tf.compat.v1.GraphKeys.TRAINABLE_VARIABLES, scope=tf.compat.v1.get_variable_scope().name + "/q2_func")
# target q network evalution
q1_tp1 = q_func(obs_tp1_input.get(), num_actions, scope="target_q1_func", reuse=False)
target_q1_func_vars = tf.compat.v1.get_collection(tf.compat.v1.GraphKeys.TRAINABLE_VARIABLES, scope=tf.compat.v1.get_variable_scope().name + "/target_q1_func")
q2_tp1 = q_func(obs_tp1_input.get(), num_actions, scope="target_q2_func", reuse=False)
target_q2_func_vars = tf.compat.v1.get_collection(tf.compat.v1.GraphKeys.TRAINABLE_VARIABLES, scope=tf.compat.v1.get_variable_scope().name + "/target_q2_func")
# q scores for actions which we know were selected in the given state.
q1_t_selected = tf.reduce_sum(input_tensor=q1_t * tf.one_hot(act_t_ph, num_actions), axis=1)
q2_t_selected = tf.reduce_sum(input_tensor=q2_t * tf.one_hot(act_t_ph, num_actions), axis=1)
# Actions selected with current q funcs at state t+1.
q1_tp1_using_online_net = q_func(obs_tp1_input.get(), num_actions, scope="q1_func", reuse=True)
q2_tp1_using_online_net = q_func(obs_tp1_input.get(), num_actions, scope="q2_func", reuse=True)
tp1_best_action_using_online_net = tf.argmax(input=q1_tp1_using_online_net+q2_tp1_using_online_net, axis=1)
# Using action at t+1 find target value associated with the action
q1_tp1_selected = tf.reduce_sum(input_tensor=q1_tp1 * tf.one_hot(tp1_best_action_using_online_net, num_actions), axis=1)
q2_tp1_selected = tf.reduce_sum(input_tensor=q2_tp1 * tf.one_hot(tp1_best_action_using_online_net, num_actions), axis=1)
# Min of target q values to be used bellman equation
q_tp1_best = tf.minimum(q1_tp1_selected, q2_tp1_selected)
# Done mask
# q_tp1_best_masked = (1.0 - done_mask_ph) * q_tp1_best
# compute RHS of bellman equation
q_tp1_selected_target = rew_t_ph + gamma * q_tp1_best
# compute the error (potentially clipped)
td_error1 = q1_t_selected - tf.stop_gradient(q_tp1_selected_target)
td_error2 = q2_t_selected - tf.stop_gradient(q_tp1_selected_target)
errors1 = U.huber_loss(td_error1)
errors2 = U.huber_loss(td_error2)
errors = errors1 + errors2
weighted_error = tf.reduce_mean(input_tensor=importance_weights_ph * errors)
#Print total number of params
total_parameters = 0
for variable in tf.compat.v1.trainable_variables():
# shape is an array of tf.Dimension
shape = variable.get_shape()
variable_parameters = 1
for dim in shape:
variable_parameters *= dim.value
# print("var params", variable_parameters)
total_parameters += variable_parameters
print("===============================================================")
print("Total number of trainable params:", total_parameters)
print("===============================================================")
# Log for tensorboard
tf.summary.scalar('q1_values', tf.math.reduce_mean(q1_t))
tf.summary.scalar('q2_values', tf.math.reduce_mean(q2_t))
tf.summary.scalar('td_1', tf.math.reduce_mean(td_error1))
tf.summary.scalar('td_2', tf.math.reduce_mean(td_error2))
tf.summary.scalar('weighted_loss', weighted_error)
tf.summary.scalar('lr_schedule', lr)
tf.summary.scalar('td_MSE_1', tf.math.reduce_mean(tf.math.square(td_error1)))
tf.summary.scalar('td_MSE_2', tf.math.reduce_mean(tf.math.square(td_error2)))
# combine variable scopes
q_func_vars = q1_func_vars+q2_func_vars
# compute optimization op (potentially with gradient clipping)
if grad_norm_clipping is not None:
gradients = optimizer.compute_gradients(weighted_error, var_list=q_func_vars)
for i, (grad, var) in enumerate(gradients):
if grad is not None:
gradients[i] = (tf.clip_by_norm(grad, grad_norm_clipping), var)
optimize_expr = optimizer.apply_gradients(gradients)
else:
optimize_expr = optimizer.minimize(weighted_error, var_list=q_func_vars)
# update_target_fn will be called every step to copy Q network to target Q network
# target network is updated with polyak averaging
update_target_expr1 = []
for var, var_target in zip(sorted(q1_func_vars, key=lambda v: v.name),
sorted(target_q1_func_vars, key=lambda v: v.name)):
update_target_expr1.append(var_target.assign(tau*var + (1-tau)*var_target))
update_target_expr1 = tf.group(*update_target_expr1)
update_target_expr2 = []
for var, var_target in zip(sorted(q2_func_vars, key=lambda v: v.name),
sorted(target_q2_func_vars, key=lambda v: v.name)):
update_target_expr2.append(var_target.assign(tau*var + (1-tau)*var_target))
update_target_expr2 = tf.group(*update_target_expr2)
merged_summary = tf.compat.v1.summary.merge_all(scope=tf.compat.v1.get_variable_scope().name)
# Create callable functions
train = U.function(
inputs=[
obs_t_input,
act_t_ph,
rew_t_ph,
obs_tp1_input,
done_mask_ph,
importance_weights_ph,
iteration
],
outputs=[td_error1, td_error2, tf.reduce_mean(input_tensor=errors), merged_summary],
updates=[optimize_expr, lr]
)
update_target = U.function([], [], updates=[update_target_expr1, update_target_expr2])
q_values = U.function(inputs=[obs_t_input], outputs=[q1_t, q2_t])
return act_f, act_greedy, q_values, train, update_target, {'q_values': q_values}
def build_train(make_obs_ph,
q_func,
num_actions,
optimizer_f,
grad_norm_clipping=None,
gamma=1.0,
double_q=False,
scope="setdeepq",
reuse=None,
param_noise=False,
param_noise_filter_func=None,
test_eps=0.05,
lr_init=0.001,
lr_decay_factor=0.99,
lr_growth_factor=1.001,
tau=0.001):
"""Creates the train function:
Parameters
----------
make_obs_ph: str -> tf.placeholder or TfInput
a function that takes a name and creates a placeholder of input with that name
q_func: (tf.Variable, int, str, bool) -> tf.Variable
the model that takes the following inputs:
observation_in: object
the output of observation placeholder
num_actions: int
number of actions
scope: str
reuse: bool
should be passed to outer variable scope
and returns a tensor of shape (batch_size, num_actions) with values of every action.
num_actions: int
number of actions
reuse: bool
whether or not to reuse the graph variables
optimizer: tf.train.Optimizer
optimizer to use for the Q-learning objective.
grad_norm_clipping: float or None
clip gradient norms to this value. If None no clipping is performed.
gamma: float
discount rate.
double_q: bool
if true will use Double Q Learning (https://arxiv.org/abs/1509.06461).
In general it is a good idea to keep it enabled.
scope: str or VariableScope
optional scope for variable_scope.
reuse: bool or None
whether or not the variables should be reused. To be able to reuse the scope must be given.
param_noise: bool
whether or not to use parameter space noise (https://arxiv.org/abs/1706.01905)
param_noise_filter_func: tf.Variable -> bool
function that decides whether or not a variable should be perturbed. Only applicable
if param_noise is True. If set to None, default_param_noise_filter is used by default.
lr_init : float
initial learning rate
lr_decay_factor : float
learning rate decay factor. It should be equal to or smaller than 1.0.
lr_growth_factor : float
learning rate growth factor. It should be equal to or larger than 1.0.
tau : float
parameter for the soft target network update. tau <= 1.0 and 1.0 for
the hard update.
Returns
-------
act: (tf.Variable, bool, float) -> tf.Variable
function to select and action given observation.
` See the top of the file for details.
train: (object, np.array, np.array, object, np.array, np.array) -> np.array
optimize the error in Bellman's equation.
` See the top of the file for details.
update_target: () -> ()
copy the parameters from optimized Q function to the target Q function.
` See the top of the file for details.
debug: {str: function}
a bunch of functions to print debug data like q_values.
"""
# Build action graphs
act_f = build_act(make_obs_ph,
q_func,
num_actions,
scope=scope,
reuse=reuse)
act_greedy = build_act_greedy(make_obs_ph,
q_func,
num_actions,
scope=scope,
reuse=True,
eps=test_eps)
with tf.compat.v1.variable_scope(scope, reuse=reuse):
# set up placeholders
obs_t_input = make_obs_ph("obs_t")
act_t_ph = tf.compat.v1.placeholder(tf.int32, [None], name="action")
rew_t_ph = tf.compat.v1.placeholder(tf.float32, [None], name="reward")
obs_tp1_input = make_obs_ph("obs_tp1")
done_mask_ph = tf.compat.v1.placeholder(tf.float32, [None],
name="done")
importance_weights_ph = tf.compat.v1.placeholder(tf.float32, [None],
name="weight")
# Learning rate adjustment
lr = tf.Variable(float(lr_init), trainable=False, dtype=tf.float32)
lr_decay_op = lr.assign(
tf.clip_by_value(lr * lr_decay_factor, 1e-5, 1e-2))
lr_growth_op = lr.assign(
tf.clip_by_value(lr * lr_growth_factor, 1e-5, 1e-2))
optimizer = optimizer_f(learning_rate=lr)
# q network evaluation
q_t = q_func.forward(obs_t_input.get(),
num_actions,
scope="q_func",
reuse=True) # reuse parameters from act
q_func_vars = tf.compat.v1.get_collection(
tf.compat.v1.GraphKeys.GLOBAL_VARIABLES,
scope=tf.compat.v1.get_variable_scope().name + "/q_func")
# target q network evalution
q_tp1 = q_func.forward(obs_tp1_input.get(),
num_actions,
scope="target_q_func",
reuse=False)
target_q_func_vars = tf.compat.v1.get_collection(
tf.compat.v1.GraphKeys.GLOBAL_VARIABLES,
scope=tf.compat.v1.get_variable_scope().name + "/target_q_func")
# q scores for actions which we know were selected in the given state.
q_t_selected = tf.reduce_sum(input_tensor=q_t *
tf.one_hot(act_t_ph, num_actions),
axis=1)
# compute estimate of best possible value starting from state at t + 1
if double_q:
q_tp1_using_online_net = q_func.forward(obs_tp1_input.get(),
num_actions,
scope="q_func",
reuse=True)
q_tp1_best_using_online_net = tf.argmax(
input=q_tp1_using_online_net, axis=1)
q_tp1_best = tf.reduce_sum(
input_tensor=q_tp1 *
tf.one_hot(q_tp1_best_using_online_net, num_actions),
axis=1)
else:
q_tp1_best = tf.reduce_max(input_tensor=q_tp1, axis=1)
q_tp1_best_masked = (1.0 - done_mask_ph) * q_tp1_best
# compute RHS of bellman equation
q_t_selected_target = rew_t_ph + gamma * q_tp1_best_masked
# compute the error (potentially clipped)
td_error = q_t_selected - tf.stop_gradient(q_t_selected_target)
errors = U.huber_loss(td_error)
weighted_error = tf.reduce_mean(input_tensor=importance_weights_ph *
errors)
# Log for tensorboard
tf.summary.scalar('q_values', tf.math.reduce_mean(q_t))
tf.summary.scalar('td_MSE',
tf.math.reduce_mean(tf.math.square(td_error)))
tf.summary.scalar('weighted_loss', weighted_error)
# compute optimization op (potentially with gradient clipping)
if grad_norm_clipping is not None:
gradients = optimizer.compute_gradients(weighted_error,
var_list=q_func_vars)
for i, (grad, var) in enumerate(gradients):
if grad is not None:
gradients[i] = (tf.clip_by_norm(grad,
grad_norm_clipping), var)
optimize_expr = optimizer.apply_gradients(gradients)
else:
optimize_expr = optimizer.minimize(weighted_error,
var_list=q_func_vars)
# update_target_fn will be called periodically to copy Q network to target Q network
update_target_expr = []
for var, var_target in zip(
sorted(q_func_vars, key=lambda v: v.name),
sorted(target_q_func_vars, key=lambda v: v.name)):
update_target_expr.append(
var_target.assign(tau * var + (1 - tau) * var_target))
update_target_expr = tf.group(*update_target_expr)
merged_summary = tf.compat.v1.summary.merge_all(
scope=tf.compat.v1.get_variable_scope().name)
# Create callable functions
train = U.function(inputs=[
obs_t_input, act_t_ph, rew_t_ph, obs_tp1_input, done_mask_ph,
importance_weights_ph
],
outputs=[
td_error,
tf.reduce_mean(input_tensor=errors),
merged_summary
],
updates=[optimize_expr])
update_target = U.function([], [], updates=[update_target_expr])
q_values = U.function(inputs=[obs_t_input], outputs=q_t)
return act_f, act_greedy, q_values, train, update_target, lr_decay_op, lr_growth_op, {
'q_values': q_values
}