莫烦python强化学习系列-DQN学习（代码）
import numpy as np
import pandas as pd
import tensorflow as tf
np.random.ed(1)
tf.t_random_ed(1)
# Deep Q Network off-policy
class DeepQNetwork:
def __init__(
lf,
n_actions,
n_features,
learning_rate=0.01,
reward_decay=0.9,
e_greedy=0.9,
replace_target_iter=300,
memory_size=500,
batch_size=32,
e_greedy_increment=None,
output_graph=Fal,
):
lf.n_actions = n_actions
lf.n_features = n_features
lf.lr = learning_rate
lf.gamma = reward_decay
lf.epsilon_max = e_greedy
<_size = memory_size
lf.batch_size = batch_size
lf.epsilon_increment = e_greedy_increment
lf.epsilon = 0 if e_greedy_increment is not None el lf.epsilon_max
# total learning step
lf.learn_step_counter = 0
# initialize zero memory [s, a, r, s_]
< = np.zeros((lf.memory_size, n_features * 2 + 2))
# consist of [target_net, evaluate_net]
lf._build_net()
#tf.get_collection(key, scope=None)
#⽤来获取⼀个名称是‘key’的集合中的所有元素，返回的是⼀个列表
t_params = tf.get_collection('target_net_params')
e_params = tf.get_collection('eval_net_params')
lf.ss = tf.Session()
if output_graph:
# $ tensorboard --logdir=logs
# tf.train.SummaryWriter soon be deprecated, u following
tf.summary.FileWriter("logs/", aph)
lf.ss.run(tf.global_variables_initializer())
def _build_net(lf):
# ------------------ build evaluate_net ------------------
lf.s = tf.placeholder(tf.float32, [None, lf.n_features], name='s')
lf.s = tf.placeholder(tf.float32, [None, lf.n_features], name='s')
# input
lf.q_target = tf.placeholder(tf.float32, [None, lf.n_actions], name='Q_target') # for calculating loss with tf.variable_scope('eval_net'):
# c_names(collections_names) are the collections to store variables
c_names, n_l1, w_initializer, b_initializer = \
['eval_net_params', tf.GraphKeys.GLOBAL_VARIABLES], 10, \
tf.random_normal_initializer(0., 0.3), tf.constant_initializer(0.1) # config of layers
# first layer. collections is ud later when assign to target net
with tf.variable_scope('l1'):
w1 = tf.get_variable('w1', [lf.n_features, n_l1], initializer=w_initializer, collections=c_names)
b1 = tf.get_variable('b1', [1, n_l1], initializer=b_initializer, collections=c_names)
l1 = lu(tf.matmul(lf.s, w1) + b1)
# cond layer. collections is ud later when assign to target net
with tf.variable_scope('l2'):
w2 = tf.get_variable('w2', [n_l1, lf.n_actions], initializer=w_initializer, collections=c_names)
b2 = tf.get_variable('b2', [1, lf.n_actions], initializer=b_initializer, collections=c_names)
lf.q_eval = tf.matmul(l1, w2) + b2 #[batch_size,lf.n_action]
with tf.variable_scope('loss'):
lf.loss = tf.reduce_mean(tf.squared_difference(lf.q_target, lf.q_eval))
with tf.variable_scope('train'):
lf._train_op = tf.train.RMSPropOptimizer(lf.lr).minimize(lf.loss)
# ------------------ build target_net ------------------
lf.s_ = tf.placeholder(tf.float32, [None, lf.n_features], name='s_') # input
with tf.variable_scope('target_net'):
# c_names(collections_names) are the collections to store variables
c_names = ['target_net_params', tf.GraphKeys.GLOBAL_VARIABLES]
# first layer. collections is ud later when assign to target net
列算式
with tf.variable_scope('l1'):
w1 = tf.get_variable('w1', [lf.n_features, n_l1], initializer=w_initializer, collections=c_names)
b1 = tf.get_variable('b1', [1, n_l1], initializer=b_initializer, collections=c_names)
l1 = lu(tf.matmul(lf.s_, w1) + b1)
# cond layer. collections is ud later when assign to target net
with tf.variable_scope('l2'):
w2 = tf.get_variable('w2', [n_l1, lf.n_actions], initializer=w_initializer, collections=c_names)
b2 = tf.get_variable('b2', [1, lf.n_actions], initializer=b_initializer, collections=c_names)
lf.q_next = tf.matmul(l1, w2) + b2
def store_transition(lf, s, a, r, s_):
if not hasattr(lf, 'memory_counter'):
<_counter = 0
transition = np.hstack((s, [a, r], s_))
# replace the old memory with new memory
index = lf.memory_counter % lf.memory_size
<[index, :] = transition
<_counter += 1
def choo_action(lf, obrvation):
# to have batch dimension when feed into tf placeholder
蒲公英功效与作用obrvation = waxis, :] #shape=(1,n_features)
if np.random.uniform() < lf.epsilon:
# forward feed the obrvation and get q value for every actions
actions_value = lf.ss.run(lf.q_eval, feed_dict={lf.s: obrvation})
action = np.argmax(actions_value)#未加axis＝，返回⼀个索引数值
el:
action = np.random.randint(0, lf.n_actions)
action = np.random.randint(0, lf.n_actions)
return action
def learn(lf):
# check to replace target parameters股票涨停是什么意思
if lf.learn_step_counter % lf.replace_target_iter == 0:
lf.ss.place_target_op)
print('\ntarget_params_replaced\n')
# sample batch memory from all memory
_counter > lf.memory_size:
sample_index = np.random._size, size=lf.batch_size)
el:
sample_index = np.random._counter, size=lf.batch_size)
batch_memory = lf.memory[sample_index, :]
q_next, q_eval = lf.ss.run(
[lf.q_next, lf.q_eval],
feed_dict={
#[s, a, r, s_]
lf.s_: batch_memory[:, -lf.n_features:], # fixed params
lf.s: batch_memory[:, :lf.n_features], # newest params
})
# change q_ q_eval's action
q_target = py()
batch_index = np.arange(lf.batch_size, dtype=np.int32)
eval_act_index = batch_memory[:, lf.n_features].astype(int) #action astype(int) 转换数组的数据类型 reward = batch_memory[:, lf.n_features + 1] #reward
q_target[batch_index, eval_act_index] = reward + lf.gamma * np.max(q_next, axis=1)
"""
For example in this batch I have 2 samples and 3 actions:
q_eval =
[[1, 2, 3],
[4, 5, 6]]
q_target = q_eval =
[[1, 2, 3],
[4, 5, 6]]
Then change q_target with the real q_target the q_eval's action.
For example in:
sample 0, I took action 0, and the max q_target value is -1;
sample 1, I took action 2, and the max q_target value is -2:
q_target =
[[-1, 2, 3],
[4, 5, -2]]
So the (q_target - q_eval) becomes:
[[(-1)-(1), 0, 0],
[0, 0, (-2)-(6)]]
We then backpropagate this the corresponding action to network,
leave other action as error=0 cau we didn't choo it.
"""
# train eval network
_, lf.cost = lf.ss.run([lf._train_op, lf.loss],
feed_dict={lf.s: batch_memory[:, :lf.n_features],
lf.q_target: q_target})
晚春意思
# increasing epsilon
# increasing epsilon
励志英文lf.epsilon = lf.epsilon + lf.epsilon_increment if lf.epsilon < lf.epsilon_max el lf.epsilon_max lf.learn_step_counter += 1
def plot_cost(lf):
import matplotlib.pyplot as plt
plt.plot(np.arange(st_his)), lf.cost_his)
plt.ylabel('Cost')
plt.xlabel('training steps')
6年级数学上册plt.show()
from maze_env import Maze
from RL_brain import DeepQNetwork
def run_maze():
step = 0
for episode in range(300):
# initial obrvation
obrvation = ()
while True:
# fresh env
给领导的生日祝福
# RL choo action bad on obrvation
action = RL.choo_action(obrvation)
# RL take action and get next obrvation and reward
obrvation_, reward, done = env.step(action)
RL.store_transition(obrvation, action, reward, obrvation_)
if (step > 200) and (step % 5 == 0):
RL.learn()
# swap obrvation
obrvation = obrvation_
# break while loop when end of this episode
if done:
break
step += 1
# end of game
print('game over')
env.destroy()
if __name__ == "__main__":
# maze game
env = Maze()
RL = DeepQNetwork(env.n_actions, env.n_features,
learning_rate=0.01,
reward_decay=0.9,
e_greedy=0.9,
replace_target_iter=200,
memory_size=2000,
# output_graph=True
)
env.after(100, run_maze)
env.mainloop()装饰板材品牌
RL.plot_cost()
