"""
Asynchronous Advantage Actor Critic (A3C) with continuous action space, Reinforcement Learning.
The Pendulum example.
View more on my tutorial page: morvanzhou.github.io/tutorials/
Using:
tensorflow 1.0
gym 0.8.0
"""
import multiprocessing
import threading
import tensorflow as tf
import numpy as np
import gym
import os
import shutil
import matplotlib.pyplot as plt
GAME = 'Pendulum-v0'
OUTPUT_GRAPH = True
LOG_DIR = './log'
N_WORKERS = multiprocessing.cpu_count()
MAX_EP_STEP = 400
MAX_GLOBAL_EP = 800
GLOBAL_NET_SCOPE = 'Global_Net'拉面人生 UPDATE_GLOBAL_ITER = 5
GAMMA = 0.9
ENTROPY_BETA = 0.01
LR_A = 0.0001 # learning rate for actor
LR_C = 0.001 # learning rate for critic
GLOBAL_RUNNING_R = []
GLOBAL_EP = 0
env = gym.make(GAME)
N_S = env.observation_space.shape[0] #number of states in state space
N_A = env.action_space.shape[0] #number of actions in action space
A_BOUND = [env.action_space.low, env.action_space.high] #bound of output action
class ACNet(object):
#This class is to define the global actor-critic and local actor-critics
def __init__(self, scope, globalAC=None):
if scope == GLOBAL_NET_SCOPE: # get global network
with tf.variable_scope(scope):
仙居杨梅节self.s = tf.placeholder(tf.float32, [None, N_S], 'S')
ANALYSISESself._build_net()
# Get parameters of the actor and critic in global network
self.a_params = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, scope=scope + '/actor') self.c_params = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, scope=scope + '/critic') else: # local net, calculate losses
with tf.variable_scope(scope):
self.s = tf.placeholder(tf.float32, [None, N_S], 'S')
self.a_his = tf.placeholder(tf.float32, [None, N_A], 'A')
self.v_target = tf.placeholder(tf.float32, [None, 1], 'Vtarget')
mu, sigma, self.v = self._build_net()
#self.v:state-value calculated by the critic
td = tf.subtract(self.v_target, self.v, name='TD_error')
td = tf.subtract(self.v_target, self.v, name='TD_error')
with tf.name_scope('c_loss'):
#to minimize TD-error
self.c_loss = tf.reduce_mean(tf.square(td))
with tf.name_scope('wrap_a_out'):
mu, sigma = mu * A_BOUND[1], sigma + 1e-4
#distribution of parameters:mu, sigma
normal_dist = tf.contrib.distributions.Normal(mu, sigma)
with tf.name_scope('a_loss'):
log_prob = normal_dist.log_prob(self.a_his) #log pi(a)
exp_v = log_prob * td
entropy = py()
# encourage exploration:larger entropy means more stochastic actions
self.a_loss = tf.reduce_mean(-p_v)
#to duce_p_v) <=> to duce_mean(-p_v)
with tf.name_scope('choose_a'): # use local params to choose action
self.A = tf.clip_by_value(tf.squeeze(normal_dist.sample(1), axis=0), A_BOUND[0], A_BOUND[1]) with tf.name_scope('local_grad'):
self.a_params = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, scope=scope + '/actor') self.c_params = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, scope=scope + '/critic') #gradients of a_ a_params
self.a_grads = tf.gradients(self.a_loss, self.a_params)
self.c_grads = tf.gradients(self.c_loss, self.c_params)
with tf.name_scope('sync'):
with tf.name_scope('pull'):
# assign params of global net to local net
self.pull_a_params_op = [l_p.assign(g_p) for l_p, g_p in zip(self.a_params, globalAC.a_params)] self.pull_c_params_op = [l_p.assign(g_p) for l_p, g_p in zip(self.c_params, globalAC.c_params)] with tf.name_scope('push'):
# update params of global net by pushing the calculated gradients of local net to global net
self.update_a_op = OPT_A.apply_gradients(zip(self.a_grads, globalAC.a_params))
self.update_c_op = OPT_C.apply_gradients(zip(self.c_grads, globalAC.c_params))
def _build_net(self ):
w_init = tf.random_normal_initializer(0., .1)
with tf.variable_scope('actor'):
l_a = tf.layers.dense(self.s, 200, lu6, kernel_initializer=w_init, name='la')
# N_A means the numbers of possible actions and the number of normal distributions.
mu = tf.layers.dense(l_a, N_A, tf.nn.tanh, kernel_initializer=w_init, name='mu')
sigma = tf.layers.dense(l_a, N_A, tf.nn.softplus, kernel_initializer=w_init, name='sigma')
with tf.variable_scope('critic'):
l_c = tf.layers.dense(self.s, 100, lu6, kernel_initializer=w_init, name='lc')
v = tf.layers.dense(l_c, 1, kernel_initializer=w_init, name='v') # state value
return mu, sigma, v
def update_global(self, feed_dict): # run by a local
SESS.run([self.update_a_op, self.update_c_op], feed_dict) # local grads applies to global net
def pull_global(self): # run by a local
SESS.run([self.pull_a_params_op, self.pull_c_params_op])
def choose_action(self, s):
# run by a local: choose action from normal distributions
s = waxis, :]
return SESS.run(self.A, {self.s: s})[0]
class Worker(object):
# push local gradients to global net and assign global params to local net
def __init__(self, name, globalAC):
self.name = name
self.name = name
self.AC = ACNet(name, globalAC)
def work(self):内外接
global GLOBAL_RUNNING_R, GLOBAL_EP
total_step = 1
buffer_s, buffer_a, buffer_r = [], [], []
while not COORD.should_stop() and GLOBAL_EP < MAX_GLOBAL_EP:
s = set()
ep_r = 0
for ep_t in range(MAX_EP_STEP):
if self.name == 'W_0':
a = self.AC.choose_action(s)
s_, r, done, info = v.step(a)
done = True if ep_t == MAX_EP_STEP - 1 else False
r /= 10 # normalize reward
ep_r += r
buffer_s.append(s)
buffer_a.append(a)年降雨量
buffer_r.append(r)
if total_step % UPDATE_GLOBAL_ITER == 0 or done: # update global and assign to local net
if done:
v_s_ = 0 # terminal
else:
v_s_ = SESS.run(self.AC.v, {self.AC.s: s_[np.newaxis, :]})[0, 0]
buffer_v_target = []
for r in buffer_r[::-1]: # reverse buffer r
v_s_ = r + GAMMA * v_s_
buffer_v_target.append(v_s_)
buffer_verse()
buffer_s, buffer_a, buffer_v_target = np.vstack(buffer_s), np.vstack(buffer_a), np.vstack(buffer_v_target) feed_dict = {
self.AC.s: buffer_s,
self.AC.a_his: buffer_a,
self.AC.v_target: buffer_v_target,
三权分立的弊端
}
self.AC.update_global(feed_dict) # push local gradients to global net
buffer_s, buffer_a, buffer_r = [], [], []
self.AC.pull_global() #pull the newest global params to local
s = s_
total_step += 1
if done:
if len(GLOBAL_RUNNING_R) == 0: # record running episode reward
GLOBAL_RUNNING_R.append(ep_r)
else:
GLOBAL_RUNNING_R.append(0.9 * GLOBAL_RUNNING_R[-1] + 0.1 * ep_r)
print(
self.name,
"Ep:", GLOBAL_EP,
"| Ep_r: %i" % GLOBAL_RUNNING_R[-1],
)
GLOBAL_EP += 1
break
if __name__ == "__main__":
SESS = tf.Session()
with tf.device("/cpu:0"):
# define two optimizers for actors and critics in local net
OPT_A = tf.train.RMSPropOptimizer(LR_A, name='RMSPropA')
OPT_C = tf.train.RMSPropOptimizer(LR_C, name='RMSPropC')
OPT_C = tf.train.RMSPropOptimizer(LR_C, name='RMSPropC')
# build the global net which does not calculate loss thus does not need optimizers. GLOBAL_AC = ACNet(GLOBAL_NET_SCOPE) # we only need its params
workers = []
# Create worker
for i in range(N_WORKERS):
i_name = 'W_%i' % i # worker name
workers.append(Worker(i_name, GLOBAL_AC))
COORD = tf.train.Coordinator()
SESS.run(tf.global_variables_initializer())
if OUTPUT_GRAPH:
if ists(LOG_DIR):
<(LOG_DIR)
tf.summary.FileWriter(LOG_DIR, aph)
worker_threads = []
for worker in workers:
job = lambda: worker.work()
t = threading.Thread(target=job)
t.start()
worker_threads.append(t)
COORD.join(worker_threads)
plt.plot(np.arange(len(GLOBAL_RUNNING_R)), GLOBAL_RUNNING_R)
plt.xlabel('step')
plt.ylabel('Total moving reward')
plt.show()