def c_train(make_obs_ph_n,
make_target_loc_ph_n,
c_index,
c_func,
q_func,
optimizer,
scope="trainer",
num_units=128,
grad_norm_clipping=None,
reuse=tf.AUTO_REUSE,
local_q_func=False):
with tf.variable_scope(scope, reuse=reuse):
# set up placeholders
obs_ph_n = make_obs_ph_n
target_loc_ph = make_target_loc_ph_n[
c_index] #tf.placeholder(tf.float32, [None,2], name="target_loc")
self_obs_ph = obs_ph_n[c_index]
labels_ph = tf.placeholder(tf.float32, [None, 2], name="labels")
# prior network
c_input = tf.concat((self_obs_ph, target_loc_ph), 1)
c = c_func(c_input, 2, scope="c_func", type='cls', num_units=num_units)
c_pred = tf.nn.softmax(c)
c_flags = tf.greater(c_pred[:, 0], 0.5)
c_func_vars = U.scope_vars(U.absolute_scope_name("c_func"))
# loss and optimization
loss = tf.reduce_mean(
tf.nn.softmax_cross_entropy_with_logits(logits=c,
labels=labels_ph))
optimize_expr = U.minimize_and_clip(optimizer, loss, c_func_vars,
grad_norm_clipping)
# Create callable functions
c_train = U.function(
inputs=[obs_ph_n[c_index], target_loc_ph, labels_ph],
outputs=loss,
updates=[optimize_expr])
c_act = U.function(inputs=[obs_ph_n[c_index], target_loc_ph],
outputs=c_flags)
c_values = U.function([obs_ph_n[c_index], target_loc_ph],
outputs=c_pred)
# target network
target_c_values = c_func(c_input,
2,
scope="target_c_func",
type='cls',
num_units=num_units)
target_c_pred = tf.nn.softmax(target_c_values)
target_c_flags = tf.greater(target_c_pred[:, 0], 0.5)
target_c_func_vars = U.scope_vars(
U.absolute_scope_name("target_c_func"))
update_target_c = make_update_exp(c_func_vars, target_c_func_vars)
target_c_act = U.function(inputs=[obs_ph_n[c_index], target_loc_ph],
outputs=target_c_flags)
return c_act, c_train, update_target_c, {
'c_values': c_values,
'target_c_act': target_c_act
}
def q_train(make_obs_ph_n,
act_space_n,
q_index,
q_func,
optimizer,
grad_norm_clipping=None,
local_q_func=False,
scope="trainer",
reuse=tf.AUTO_REUSE,
num_units=128):
with tf.variable_scope(scope, reuse=reuse):
# create distribtuions
act_pdtype_n = [make_pdtype(act_space) for act_space in act_space_n]
# set up placeholders
obs_ph_n = make_obs_ph_n
act_ph_n = [
act_pdtype_n[i].sample_placeholder([None], name="action" + str(i))
for i in range(len(act_space_n))
]
target_ph = tf.placeholder(tf.float32, [None], name="target")
# q network
q_input = tf.concat(obs_ph_n + act_ph_n, 1)
if local_q_func:
q_input = tf.concat([obs_ph_n[q_index], act_ph_n[q_index]], 1)
q = q_func(q_input, 1, scope="q_func", type='fit',
num_units=num_units)[:, 0]
q_func_vars = U.scope_vars(U.absolute_scope_name("q_func"))
q_loss = tf.reduce_mean(tf.square(q - target_ph))
# viscosity solution to Bellman differential equation in place of an initial condition
q_reg = tf.reduce_mean(tf.square(q))
loss = q_loss #+ 1e-3 * q_reg
optimize_expr = U.minimize_and_clip(optimizer, loss, q_func_vars,
grad_norm_clipping)
# Create callable functions
train = U.function(inputs=obs_ph_n + act_ph_n + [target_ph],
outputs=loss,
updates=[optimize_expr])
q_values = U.function(obs_ph_n + act_ph_n, q)
# target network
target_q = q_func(q_input,
1,
scope="target_q_func",
type='fit',
num_units=num_units)[:, 0]
target_q_func_vars = U.scope_vars(
U.absolute_scope_name("target_q_func"))
update_target_q = make_update_exp(q_func_vars, target_q_func_vars)
target_q_values = U.function(obs_ph_n + act_ph_n, target_q)
return train, update_target_q, {
'q_values': q_values,
'target_q_values': target_q_values
}
def make_update_exp(vals, target_vals):
polyak = 1.0 - 1e-2
expression = []
for var, var_target in zip(sorted(vals, key=lambda v: v.name),
sorted(target_vals, key=lambda v: v.name)):
expression.append(
var_target.assign(polyak * var_target + (1.0 - polyak) * var))
expression = tf.group(*expression)
return U.function([], [], updates=[expression])
def p_m_train(make_obs_ph_n,
make_message_ph_n,
act_space_n,
num_agents_obs,
p_index,
m_func,
p_func,
q_func,
optimizer,
grad_norm_clipping=None,
local_q_func=False,
num_units=128,
scope="trainer",
reuse=tf.AUTO_REUSE):
with tf.variable_scope(scope, reuse=reuse):
# create distribtuions
act_pdtype_n = [make_pdtype(act_space) for act_space in act_space_n]
# set up placeholders
obs_ph_n = make_obs_ph_n
message_ph_n = make_message_ph_n
act_ph_n = [
act_pdtype_n[i].sample_placeholder([None], name="action" + str(i))
for i in range(len(act_space_n))
]
blz_distribution = tf.placeholder(tf.float32,
[None, act_space_n[p_index].n],
name="blz_distribution")
m_input = message_ph_n[p_index]
encode_dim = m_input.get_shape().as_list()[-1]
# message encoder
message_encode = m_func(m_input,
encode_dim,
num_agents_obs,
scope='m_func',
num_units=num_units)
m_func_vars = U.scope_vars(U.absolute_scope_name("m_func"))
# policy
p_input = tf.concat((obs_ph_n[p_index], message_encode), 1)
p = p_func(p_input,
int(act_pdtype_n[p_index].param_shape()[0]),
scope="p_func",
type='fit',
num_units=num_units)
p_func_vars = U.scope_vars(U.absolute_scope_name("p_func"))
# wrap parameters in distribution
act_pd = act_pdtype_n[p_index].pdfromflat(p)
act_sample = act_pd.sample()
p_reg = tf.reduce_mean(tf.square(act_pd.flatparam()))
act_input_n = act_ph_n + []
# correlation reg
k = tf.keras.losses.KLDivergence()
#KL_reg = k(blz_distribution, act_sample)
# q network
act_input_n[p_index] = act_pd.sample()
q_input = tf.concat(obs_ph_n + act_input_n, 1)
if local_q_func:
q_input = tf.concat([obs_ph_n[p_index], act_input_n[p_index]], 1)
q = q_func(q_input,
1,
scope="q_func",
type='fit',
reuse=True,
num_units=num_units)[:, 0]
# loss and optimization
pg_loss = -tf.reduce_mean(q)
loss = pg_loss #+ KL_reg * 1e-2
optimize_expr = U.minimize_and_clip(optimizer, loss,
[p_func_vars, m_func_vars],
grad_norm_clipping)
# Create callable functions
train = U.function(inputs=obs_ph_n + message_ph_n + act_ph_n +
[blz_distribution],
outputs=loss,
updates=[optimize_expr])
act = U.function(inputs=[obs_ph_n[p_index], message_ph_n[p_index]],
outputs=act_sample)
p_values = U.function([obs_ph_n[p_index], message_ph_n[p_index]],
outputs=p)
# target network
target_message_encode = m_func(m_input,
encode_dim,
num_agents_obs,
scope='target_m_func',
num_units=num_units)
target_m_func_vars = U.scope_vars(
U.absolute_scope_name("target_m_func"))
p_input = tf.concat((obs_ph_n[p_index], target_message_encode), 1)
target_p = p_func(p_input,
int(act_pdtype_n[p_index].param_shape()[0]),
scope="target_p_func",
type='fit',
num_units=num_units)
target_p_func_vars = U.scope_vars(
U.absolute_scope_name("target_p_func"))
update_target_m = make_update_exp(m_func_vars, target_m_func_vars)
update_target_p = make_update_exp(p_func_vars, target_p_func_vars)
target_act_sample = act_pdtype_n[p_index].pdfromflat(target_p).sample()
target_act = U.function(
inputs=[obs_ph_n[p_index], message_ph_n[p_index]],
outputs=target_act_sample)
return act, train, update_target_p, update_target_m, {
'p_values': p_values,
'target_act': target_act
}