def p_act(make_obs_ph_n,
act_space_n,
p_index,
p_func,
lstm_model,
num_units=64,
scope="trainer",
reuse=None,
use_lstm=True):
with tf.variable_scope(scope, reuse=reuse):
# ============p network建图=================
# batch size的placeholder, []
batch_size = tf.placeholder(tf.int32, shape=[], name="bs")
# observation placeholder, list of [batch_size, dim, time_step]
obs_ph_n = make_obs_ph_n
# action distribution
act_pdtype_n = [make_pdtype(act_space)
for act_space in act_space_n] # 创建action的分布用来采样
if use_lstm:
observation_n = lstm_model(obs_ph_n, reuse=reuse, scope="lstm")
else:
observation_n = [tf.squeeze(o, 2) for o in obs_ph_n]
# 当前智能体的局部obs, [batch_size, state_dim]
p_input = observation_n[p_index]
# 计算局部p值，最后用来产生action, [batch_size, action_dim]
p = p_func(p_input,
int(act_pdtype_n[p_index].param_shape()[0]),
scope="p_func",
num_units=num_units,
reuse=reuse)
# p_func_vars = U.scope_vars(U.absolute_scope_name("p_func"))
act_pd = act_pdtype_n[p_index].pdfromflat(p) #
act_sample = act_pd.sample() # [batch_size, action_dim]
# ============p network建图结束=================
# 采样aciton的函数调用
act = U.function(inputs=[obs_ph_n[p_index]] + [batch_size],
outputs=act_sample)
# p value的函数调用
p_values = U.function([obs_ph_n[p_index]] + [batch_size], p)
# ============target p network建图=================
target_p = p_func(p_input,
int(act_pdtype_n[p_index].param_shape()[0]),
scope="target_p_func",
reuse=reuse,
num_units=num_units)
# target_p_func_vars = U.scope_vars(U.absolute_scope_name("target_p_func"))
target_act_pd = act_pdtype_n[p_index].pdfromflat(
target_p) # target action, [batch_size, action_dim]
target_act_sample = target_act_pd.sample()
# ============p target network建图结束=================
# target action的函数调用
target_act = U.function(inputs=[obs_ph_n[p_index]] + [batch_size],
outputs=target_act_sample)
return act, {'p_values': p_values, 'target_act': target_act}
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 q_train(make_obs_ph_n,
act_space_n,
q_index,
q_func,
lstm_model,
optimizer,
args,
grad_norm_clipping=None,
local_q_func=False,
scope="trainer",
reuse=None,
num_units=64,
use_lstm=True):
with tf.variable_scope(scope, reuse=reuse):
# ===================q network开始建图=================
batch_size = tf.placeholder(tf.int32, shape=[], name="bs")
# 创建分布
act_pdtype_n = [make_pdtype(act_space) for act_space in act_space_n]
# 创建观测placeholder
obs_ph_n = make_obs_ph_n # set up placeholders
# 在这里进行dimension reduction
if use_lstm:
observation_n = lstm_model(obs_ph_n, scope="lstm", reuse=reuse)
else:
observation_n = [tf.squeeze(o, 2) for o in 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") # 在运行时计算，然后传入，只跟loss有关
# 所有智能体的obs和action
if local_q_func:
q_input = tf.concat([observation_n[q_index], act_ph_n[q_index]], 1)
else:
q_input = tf.concat(observation_n + act_ph_n, 1)
q = q_func(q_input, 1, scope="q_func", num_units=num_units)[:,
0] # 计算q值
q_func_vars = U.scope_vars(
U.absolute_scope_name("q_func")) # q network网络参数
if use_lstm:
lstm_func_vars = U.scope_vars(
U.absolute_scope_name("lstm")) # lstm参数
q_loss = tf.reduce_mean(tf.square(q - target_ph)) # mse loss
q_reg = tf.reduce_mean(tf.square(q))
loss = q_loss # + 1e-3 * q_reg
if use_lstm:
optimize_expr = U.minimize_and_clip(optimizer, loss,
q_func_vars + lstm_func_vars,
grad_norm_clipping)
else:
optimize_expr = U.minimize_and_clip(optimizer, loss, q_func_vars,
grad_norm_clipping)
# ===============q network建图结束=====================
# 创建可调用函数
train = U.function(inputs=obs_ph_n + act_ph_n + [target_ph] +
[batch_size],
outputs=loss,
updates=[optimize_expr])
q_values = U.function(obs_ph_n + act_ph_n + [batch_size], q)
# ==================target q network建图===============
target_q = q_func(q_input,
1,
scope="target_q_func",
num_units=num_units)[:, 0]
target_q_func_vars = U.scope_vars(
U.absolute_scope_name("target_q_func"))
# ===================target q network建图结束======================
# 创建可调用函数
update_target_q = make_update_exp(q_func_vars, target_q_func_vars)
target_q_values = U.function(obs_ph_n + act_ph_n + [batch_size],
target_q)
return train, update_target_q, {
'q_values': q_values,
'target_q_values': target_q_values
}
def p_train(make_obs_ph_n,
act_space_n,
p_scope,
p_index,
p_func,
q_func,
lstm_model,
optimizer,
args,
grad_norm_clipping=None,
local_q_func=False,
num_units=64,
scope="trainer",
reuse=None,
use_lstm=True):
with tf.variable_scope(scope, reuse=reuse):
# batch size的placeholder, []
batch_size = tf.placeholder(tf.int32, shape=[], name="bs")
# 创建action的分布用来采样
act_pdtype_n = [make_pdtype(act_space) for act_space in act_space_n]
# 创建observation的placeholder # list of [batch_size, dim, time_step]
obs_ph_n = make_obs_ph_n
# action的placeholder, list of [batch_size, action_dim
act_ph_n = [
act_pdtype_n[i].sample_placeholder([None], name="action" + str(i))
for i in range(len(act_space_n))
]
# 所有智能体的obs, list of [batch_size, state_dim]
if use_lstm:
observation_n = lstm_model(obs_ph_n, reuse=reuse, scope="lstm")
else:
observation_n = [tf.squeeze(o, 2) for o in obs_ph_n]
# 当前智能体的局部obs, [batch_size, state_dim]
p_input = observation_n[p_index]
# 计算局部p值，最后用来产生action, [batch_size, action_dim]
p = p_func(p_input,
int(act_pdtype_n[p_index].param_shape()[0]),
scope="p_func",
reuse=reuse,
num_units=num_units)
# p函数的参数
p_func_vars = U.scope_vars(U.absolute_scope_name("p_func"))
if use_lstm:
lstm_vars = U.scope_vars(U.absolute_scope_name("lstm"))
# wrap parameters in distribution,Pd.logits
act_pd = act_pdtype_n[p_index].pdfromflat(p) #
# act_sample = act_pd.sample() # [batch_size, action_dim]
p_reg = tf.reduce_mean(tf.square(act_pd.flatparam())) # [None]
# 更新action
act_input_n = act_ph_n + []
act_input_n[p_index] = act_pd.sample() # [batch_size, action]
# 所有智能体的s和a, [batch_size, concat_dim]
q_input = tf.concat(observation_n + act_input_n, 1)
if local_q_func:
q_input = tf.concat([obs_ph_n[p_index], act_input_n[p_index]], 1)
# 计算Q(s,a), [batch_size,]
with tf.variable_scope(p_scope, reuse=reuse):
q = q_func(q_input,
1,
scope="q_func",
reuse=reuse,
num_units=num_units)[:, 0]
with tf.variable_scope(scope + "_" + p_scope, reuse=False):
pg_loss = -tf.reduce_mean(q) # policy gradient loss ???
loss = pg_loss + p_reg * 1e-3 # 使用每一个critic计算的loss都是不同的，第一次需要建图，以后就不需要了
# p网络的优化器。
if use_lstm:
optimize_expr = U.minimize_and_clip(optimizer, loss,
p_func_vars + lstm_vars,
grad_norm_clipping)
else:
optimize_expr = U.minimize_and_clip(optimizer, loss, p_func_vars,
grad_norm_clipping)
# ============p network建图结束=================
# 创建可以调用的函数，就是往里面喂数据
# train的调用函数，输入必须是list，
train = U.function(inputs=obs_ph_n + act_ph_n + [batch_size],
outputs=loss,
updates=[optimize_expr])
return train
def p_train(make_obs_ph_n,
act_space_n,
p_scope,
p_index,
p_func,
q_func,
lstm_model,
optimizer,
grad_norm_clipping=None,
local_q_func=False,
num_units=64,
scope="trainer",
reuse=None,
use_lstm=False):
with tf.variable_scope(scope, reuse=reuse):
# set up placeholders
# batch size的placeholder, []
batch_size = tf.placeholder(tf.int32, shape=[], name="bs")
# action placeholder
act_pdtype_n = [make_pdtype(act_space)
for act_space in act_space_n] # create distribtuions
act_ph_n = [
act_pdtype_n[i].sample_placeholder([None], name="action" + str(i))
for i in range(len(act_space_n))
]
# observation placeholder
obs_ph_n = make_obs_ph_n
if use_lstm:
observation_n = lstm_model(obs_ph_n, reuse=reuse, scope="lstm")
else:
observation_n = [tf.squeeze(o, 2) for o in obs_ph_n]
p_input = observation_n[p_index]
with tf.variable_scope(p_scope, reuse=tf.AUTO_REUSE):
p = p_func(p_input,
int(act_pdtype_n[p_index].param_shape()[0]),
scope="p_func",
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 + []
act_input_n[p_index] = act_pd.sample() # act_pd.mode() #
# target network
target_p = p_func(p_input,
int(act_pdtype_n[p_index].param_shape()[0]),
scope="target_p_func",
num_units=num_units)
target_p_func_vars = U.scope_vars(
U.absolute_scope_name("target_p_func"))
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()
with tf.variable_scope(scope, reuse=reuse):
q_input = tf.concat(observation_n + act_input_n,
1) # 所有智能体的s和a, [batch_size, concat_dim]
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", reuse=True,
num_units=num_units)[:, 0]
pg_loss = -tf.reduce_mean(q)
loss = pg_loss + p_reg * 1e-3
optimize_expr = U.minimize_and_clip(optimizer, loss, p_func_vars,
grad_norm_clipping)
# Create callable functions
train = U.function(inputs=obs_ph_n + act_ph_n + [batch_size],
outputs=loss,
updates=[optimize_expr])
act = U.function(inputs=[obs_ph_n[p_index]] + [batch_size],
outputs=act_sample)
p_values = U.function([obs_ph_n[p_index]] + [batch_size], p)
target_act = U.function(inputs=[obs_ph_n[p_index]] + [batch_size],
outputs=target_act_sample)
return act, train, update_target_p, {
'p_values': p_values,
'target_act': target_act
}
def q_train(make_obs_ph_n,
act_space_n,
q_index,
q_func,
lstm_model,
optimizer,
grad_norm_clipping=None,
local_q_func=False,
scope="trainer",
reuse=None,
num_units=64,
use_lstm=False):
with tf.variable_scope(scope, reuse=reuse):
# set up placeholders
# batch size placeholder
batch_size = tf.placeholder(tf.int32, shape=[], name="bs")
# action placeholder
act_pdtype_n = [make_pdtype(act_space)
for act_space in act_space_n] # create distribtuions
act_ph_n = [
act_pdtype_n[i].sample_placeholder([None], name="action" + str(i))
for i in range(len(act_space_n))
]
# observation placeholder
obs_ph_n = make_obs_ph_n
# target q placeholder
target_ph = tf.placeholder(tf.float32, [None], name="target")
if use_lstm:
observation_n = lstm_model(obs_ph_n, scope="lstm", reuse=reuse)
else:
observation_n = [tf.squeeze(o, 2) for o in obs_ph_n]
if local_q_func:
q_input = tf.concat([observation_n[q_index], act_ph_n[q_index]], 1)
else:
q_input = tf.concat(observation_n + act_ph_n, 1)
q = q_func(q_input, 1, scope="q_func", num_units=num_units)[:, 0]
q_func_vars = U.scope_vars(U.absolute_scope_name("q_func"))
if use_lstm:
lstm_func_vars = U.scope_vars(
U.absolute_scope_name("lstm")) # lstm参数
q_loss = tf.reduce_mean(tf.square(q - target_ph))
q_reg = tf.reduce_mean(tf.square(q))
loss = q_loss # + 1e-3 * q_reg
if use_lstm:
optimize_expr = U.minimize_and_clip(optimizer, loss,
q_func_vars + lstm_func_vars,
grad_norm_clipping)
else:
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] +
[batch_size],
outputs=loss,
updates=[optimize_expr])
q_values = U.function(obs_ph_n + act_ph_n + [batch_size], q)
# target network
target_q = q_func(q_input,
1,
scope="target_q_func",
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 + [batch_size],
target_q)
return train, update_target_q, {
'q_values': q_values,
'target_q_values': target_q_values
}
def p_act(make_obs_ph_n,
act_space_n,
p_index,
p_func,
lstm_model,
args,
num_units=64,
scope="trainer",
reuse=None):
with tf.variable_scope(scope, reuse=reuse):
# ============p network建图=================
# batch size的placeholder, []
batch_size = tf.placeholder(tf.int32, shape=[], name="bs")
# 创建action的分布用来采样
act_pdtype_n = [make_pdtype(act_space) for act_space in act_space_n]
# 创建observation的placeholder # list of [batch_size, dim, time_step]
obs_ph_n = make_obs_ph_n
# 所有智能体的obs, list of [batch_size, state_dim]
observation_n = lstm_model(obs_ph_n, reuse=reuse, scope="lstm")
# 当前智能体的局部obs, [batch_size, state_dim]
p_input = observation_n[p_index]
# 计算局部p值，最后用来产生action, [batch_size, action_dim]
p = p_func(p_input,
int(act_pdtype_n[p_index].param_shape()[0]),
scope="p_func",
num_units=num_units,
reuse=reuse)
p_func_vars = U.scope_vars(U.absolute_scope_name("p_func"))
act_pd = act_pdtype_n[p_index].pdfromflat(p) #
act_sample = act_pd.sample() # [batch_size, action_dim]
# ============p network建图结束=================
# 调用函数
# 采样aciton的调用函数
act = U.function(inputs=[obs_ph_n[p_index]] + [batch_size],
outputs=act_sample)
# 计算p值的调用函数
p_values = U.function([obs_ph_n[p_index]] + [batch_size], p)
# ============target p network建图=================
target_p = p_func(p_input,
int(act_pdtype_n[p_index].param_shape()[0]),
scope="target_p_func",
reuse=reuse,
num_units=num_units)
target_p_func_vars = U.scope_vars(
U.absolute_scope_name("target_p_func"))
# 更新目标网络的参数
# target action, [batch_size, action_dim]
target_act_pd = act_pdtype_n[p_index].pdfromflat(target_p)
target_act_sample = target_act_pd.sample()
# ============p target network建图结束=================
# 生成调用函数
# 生成目标action的调用函数
target_act = U.function(inputs=[obs_ph_n[p_index]] + [batch_size],
outputs=target_act_sample)
update_target_p = make_update_exp(p_func_vars, target_p_func_vars)
return act, update_target_p, {
'p_values': p_values,
'target_act': target_act
}