Article directory
foreword
Reinforcement learning is a large category of machine learning. It allows the machine to learn how to get high scores in the environment and perform excellent results. Behind these results is his hard work, constant trial and error, and continuous improvement. Experiment, accumulate experience, learn from experience.
Reinforcement learning methods can be divided into those who do not understand the environment. If you don't understand the environment, you can do what the environment gives you. It is called model-free, and it includes methods such as Q learning, Sarsa, and Policy Gradients. Understanding the environment and using one more model to represent the environment is the model-based method. The OpenAI gym environment library is a library that has written a variety of interactive environments, and writing the environment by yourself is a very time-consuming process. The following does not involve the writing of the environment.
1. Q learning
Q learning is a model-free method. Its core is to construct a Q table, which represents the reward value of each action (action) in each state (state). For example (don't bother with the python example), the following figure is a process of reinforcement learning, with 16 states (positions) and 4 optional actions (up, down, left, and right). Let the explorer (red box) learn to walk the maze. The yellow one is heaven (reward 1), and the black one is hell (reward -1).
Then, the process of Q learning is as follows.
Contains three steps that are repeated continuously.
- Given a current state s and a Q table, take an action a using a greedy algorithm
- Given a current state s and an action a, the next state s' and reward r are given by the environment interaction
- From the s, s', a, Q tables, the new Q table is updated to get a new Q table.
We use the Q reality and Q estimation every time we update, and the fascinating thing about Q learning is that in the Q(s1, a2) reality, it also includes A maximum estimate of Q(s2) is established, and the maximum estimate of the decay of the next step and the current reward are taken as the reality of this step .
code show as below:
import numpy as np
import pandas as pd
class QLearningTable:
def __init__(self, actions, learning_rate=0.01, reward_decay=0.9, e_greedy=0.9):
self.actions = actions # a list
self.lr = learning_rate
self.gamma = reward_decay
self.epsilon = e_greedy
self.q_table = pd.DataFrame(columns=self.actions, dtype=np.float64)
def choose_action(self, observation):
self.check_state_exist(observation)
# action selection
if np.random.uniform() < self.epsilon:
# choose best action
state_action = self.q_table.loc[observation, :]
# some actions may have the same value, randomly choose on in these actions
action = np.random.choice(state_action[state_action == np.max(state_action)].index)
else:
# choose random action
action = np.random.choice(self.actions)
return action
def learn(self, s, a, r, s_):
self.check_state_exist(s_)
q_predict = self.q_table.loc[s, a]
if s_ != 'terminal':
q_target = r + self.gamma * self.q_table.loc[s_, :].max() # next state is not terminal
else:
q_target = r # next state is terminal
self.q_table.loc[s, a] += self.lr * (q_target - q_predict) # update
def check_state_exist(self, state):
if state not in self.q_table.index:
# append new state to q table
self.q_table = self.q_table.append(
pd.Series(
[0]*len(self.actions),
index=self.q_table.columns,
name=state,
)
)
from maze_env import Maze
from RL_brain import QLearningTable
def update():
for episode in range(100):
# initial observation
observation = env.reset()
while True:
# fresh env
env.render()
# RL choose action based on observation
action = RL.choose_action(str(observation))
# RL take action and get next observation and reward
observation_, reward, done = env.step(action)
# RL learn from this transition
RL.learn(str(observation), action, reward, str(observation_))
# swap observation
observation = observation_
# break while loop when end of this episode
if done:
break
# end of game
print('game over')
env.destroy()
if __name__ == "__main__":
env = Maze()
RL = QLearningTable(actions=list(range(env.n_actions)))
env.after(100, update)
env.mainloop()
2. Sauce
Sarsa is very similar to Q learning, the difference is that Sarsa will be more 'timid' and less daring to try. Its flow is as follows.
It can be seen that the difference between it and Q learning is only in the update link, specifically:
- He has already thought about the action corresponding to the state in the current state, and he has thought about the next state_ and the next action_ (Qlearning has not yet thought about the next action_)
- When updating Q(s,a), it is based on the Q(s_, a_) of the next greedy algorithm (Qlearning is based on maxQ(s_)). This difference makes Sarsa more timid than Qlearning. Because
Qlearning Always think about maxQ maximization, because of this maxQ, it becomes greedy, regardless of other non-maxQ results. We can understand that Qlearning is a greedy, bold, and brave algorithm, and it doesn’t care about mistakes and death. And Sarsa It is a conservative algorithm, he cares about every step of decision-making, and is sensitive to mistakes and death.
import numpy as np
import pandas as pd
class RL(object):
def __init__(self, action_space, learning_rate=0.01, reward_decay=0.9, e_greedy=0.9):
self.actions = action_space # a list
self.lr = learning_rate
self.gamma = reward_decay
self.epsilon = e_greedy
self.q_table = pd.DataFrame(columns=self.actions, dtype=np.float64)
def check_state_exist(self, state):
if state not in self.q_table.index:
# append new state to q table
self.q_table = self.q_table.append(
pd.Series(
[0]*len(self.actions),
index=self.q_table.columns,
name=state,
)
)
def choose_action(self, observation):
self.check_state_exist(observation)
# action selection
if np.random.rand() < self.epsilon:
# choose best action
state_action = self.q_table.loc[observation, :]
# some actions may have the same value, randomly choose on in these actions
action = np.random.choice(state_action[state_action == np.max(state_action)].index)
else:
# choose random action
action = np.random.choice(self.actions)
return action
def learn(self, *args):
pass
# off-policy
class QLearningTable(RL):
def __init__(self, actions, learning_rate=0.01, reward_decay=0.9, e_greedy=0.9):
super(QLearningTable, self).__init__(actions, learning_rate, reward_decay, e_greedy)
def learn(self, s, a, r, s_):
self.check_state_exist(s_)
q_predict = self.q_table.loc[s, a]
if s_ != 'terminal':
q_target = r + self.gamma * self.q_table.loc[s_, :].max() # next state is not terminal
else:
q_target = r # next state is terminal
self.q_table.loc[s, a] += self.lr * (q_target - q_predict) # update
# on-policy
class SarsaTable(RL):
def __init__(self, actions, learning_rate=0.01, reward_decay=0.9, e_greedy=0.9):
super(SarsaTable, self).__init__(actions, learning_rate, reward_decay, e_greedy)
def learn(self, s, a, r, s_, a_):
self.check_state_exist(s_)
q_predict = self.q_table.loc[s, a]
if s_ != 'terminal':
q_target = r + self.gamma * self.q_table.loc[s_, a_] # next state is not terminal
else:
q_target = r # next state is terminal
self.q_table.loc[s, a] += self.lr * (q_target - q_predict) # update
from maze_env import Maze
from RL_brain import SarsaTable
def update():
for episode in range(100):
# 初始化环境
observation = env.reset()
# Sarsa 根据 state 观测选择行为
action = RL.choose_action(str(observation))
while True:
# 刷新环境
env.render()
# 在环境中采取行为, 获得下一个 state_ (obervation_), reward, 和是否终止
observation_, reward, done = env.step(action)
# 根据下一个 state (obervation_) 选取下一个 action_
action_ = RL.choose_action(str(observation_))
# 从 (s, a, r, s, a) 中学习, 更新 Q_tabel 的参数 ==> Sarsa
RL.learn(str(observation), action, reward, str(observation_), action_)
# 将下一个当成下一步的 state (observation) and action
observation = observation_
action = action_
# 终止时跳出循环
if done:
break
# 大循环完毕
print('game over')
env.destroy()
if __name__ == "__main__":
env = Maze()
RL = SarsaTable(actions=list(range(env.n_actions)))
env.after(100, update)
env.mainloop()
3. Deep Q Network(DQN)
DQN is a type of reinforcement learning combined with neural networks. In ordinary reinforcement learning, a Q table needs to be generated, and if there are too many states, the Q table will consume a lot of memory, so DQN proposes the function of replacing the Q table with a neural network. The network inputs a state and outputs the Q value of each action. The network updates parameters by using RMSprop on the Q-estimate and Q-reality. The Q estimate is the network output, and the Q reality is equal to the reward + the previous model 's Q estimate for the next state. The flow chart is as follows:
the whole algorithm looks complicated at first glance, but it becomes simpler when we split it up. That is to say, some decorations are added to the main frame of Q learning, including:
- memory bank (for repeated learning)
- The neural network calculates the Q value
- Temporarily freeze the q_target parameter (cut dependencies)
Specifically, the memory bank stores a bunch of data in a constantly updated memory bank, and randomly extracts data for training during training. The neural network is used to output the Q value of each action for the input state. Two networks are shared, and their structures are exactly the same, but the q_target network uses the parameters of many steps before the main network, which is to form a delay and cut off their correlation.
import numpy as np
import pandas as pd
import tensorflow as tf
np.random.seed(1)
tf.set_random_seed(1)
# Deep Q Network off-policy
class DeepQNetwork:
def __init__(
self,
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=False,
):
self.n_actions = n_actions
self.n_features = n_features
self.lr = learning_rate
self.gamma = reward_decay
self.epsilon_max = e_greedy
self.replace_target_iter = replace_target_iter
self.memory_size = memory_size
self.batch_size = batch_size
self.epsilon_increment = e_greedy_increment
self.epsilon = 0 if e_greedy_increment is not None else self.epsilon_max
# total learning step
self.learn_step_counter = 0
# initialize zero memory [s, a, r, s_]
self.memory = np.zeros((self.memory_size, n_features * 2 + 2))
# consist of [target_net, evaluate_net]
self._build_net()
t_params = tf.get_collection('target_net_params')
e_params = tf.get_collection('eval_net_params')
self.replace_target_op = [tf.assign(t, e) for t, e in zip(t_params, e_params)]
self.sess = tf.Session()
if output_graph:
# $ tensorboard --logdir=logs
# tf.train.SummaryWriter soon be deprecated, use following
tf.summary.FileWriter("logs/", self.sess.graph)
self.sess.run(tf.global_variables_initializer())
self.cost_his = []
def _build_net(self):
# ------------------ build evaluate_net ------------------
self.s = tf.placeholder(tf.float32, [None, self.n_features], name='s') # input
self.q_target = tf.placeholder(tf.float32, [None, self.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 used later when assign to target net
with tf.variable_scope('l1'):
w1 = tf.get_variable('w1', [self.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 = tf.nn.relu(tf.matmul(self.s, w1) + b1)
# second layer. collections is used later when assign to target net
with tf.variable_scope('l2'):
w2 = tf.get_variable('w2', [n_l1, self.n_actions], initializer=w_initializer, collections=c_names)
b2 = tf.get_variable('b2', [1, self.n_actions], initializer=b_initializer, collections=c_names)
self.q_eval = tf.matmul(l1, w2) + b2
with tf.variable_scope('loss'):
self.loss = tf.reduce_mean(tf.squared_difference(self.q_target, self.q_eval))
with tf.variable_scope('train'):
self._train_op = tf.train.RMSPropOptimizer(self.lr).minimize(self.loss)
# ------------------ build target_net ------------------
self.s_ = tf.placeholder(tf.float32, [None, self.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 used later when assign to target net
with tf.variable_scope('l1'):
w1 = tf.get_variable('w1', [self.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 = tf.nn.relu(tf.matmul(self.s_, w1) + b1)
# second layer. collections is used later when assign to target net
with tf.variable_scope('l2'):
w2 = tf.get_variable('w2', [n_l1, self.n_actions], initializer=w_initializer, collections=c_names)
b2 = tf.get_variable('b2', [1, self.n_actions], initializer=b_initializer, collections=c_names)
self.q_next = tf.matmul(l1, w2) + b2
def store_transition(self, s, a, r, s_):
if not hasattr(self, 'memory_counter'):
self.memory_counter = 0
transition = np.hstack((s, [a, r], s_))
# replace the old memory with new memory
index = self.memory_counter % self.memory_size
self.memory[index, :] = transition
self.memory_counter += 1
def choose_action(self, observation):
# to have batch dimension when feed into tf placeholder
observation = observation[np.newaxis, :]
if np.random.uniform() < self.epsilon:
# forward feed the observation and get q value for every actions
actions_value = self.sess.run(self.q_eval, feed_dict={
self.s: observation})
action = np.argmax(actions_value)
else:
action = np.random.randint(0, self.n_actions)
return action
def learn(self):
# check to replace target parameters
if self.learn_step_counter % self.replace_target_iter == 0:
self.sess.run(self.replace_target_op)
print('\ntarget_params_replaced\n')
# sample batch memory from all memory
if self.memory_counter > self.memory_size:
sample_index = np.random.choice(self.memory_size, size=self.batch_size)
else:
sample_index = np.random.choice(self.memory_counter, size=self.batch_size)
batch_memory = self.memory[sample_index, :]
q_next, q_eval = self.sess.run(
[self.q_next, self.q_eval],
feed_dict={
self.s_: batch_memory[:, -self.n_features:], # fixed params
self.s: batch_memory[:, :self.n_features], # newest params
})
# change q_target w.r.t q_eval's action
q_target = q_eval.copy()
batch_index = np.arange(self.batch_size, dtype=np.int32)
eval_act_index = batch_memory[:, self.n_features].astype(int)
reward = batch_memory[:, self.n_features + 1]
q_target[batch_index, eval_act_index] = reward + self.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 value w.r.t 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 error w.r.t the corresponding action to network,
leave other action as error=0 cause we didn't choose it.
"""
# train eval network
_, self.cost = self.sess.run([self._train_op, self.loss],
feed_dict={
self.s: batch_memory[:, :self.n_features],
self.q_target: q_target})
self.cost_his.append(self.cost)
# increasing epsilon
self.epsilon = self.epsilon + self.epsilon_increment if self.epsilon < self.epsilon_max else self.epsilon_max
self.learn_step_counter += 1
def plot_cost(self):
import matplotlib.pyplot as plt
plt.plot(np.arange(len(self.cost_his)), self.cost_his)
plt.ylabel('Cost')
plt.xlabel('training steps')
plt.show()
from maze_env import Maze
from RL_brain import DeepQNetwork
def run_maze():
step = 0 # 用来控制什么时候学习
for episode in range(300):
# 初始化环境
observation = env.reset()
while True:
# 刷新环境
env.render()
# DQN 根据观测值选择行为
action = RL.choose_action(observation)
# 环境根据行为给出下一个 state, reward, 是否终止
observation_, reward, done = env.step(action)
# DQN 存储记忆
RL.store_transition(observation, action, reward, observation_)
# 控制学习起始时间和频率 (先累积一些记忆再开始学习)
if (step > 200) and (step % 5 == 0):
RL.learn()
# 将下一个 state_ 变为 下次循环的 state
observation = observation_
# 如果终止, 就跳出循环
if done:
break
step += 1 # 总步数
# end of game
print('game over')
env.destroy()
if __name__ == "__main__":
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, # 每 200 步替换一次 target_net 的参数
memory_size=2000, # 记忆上限
# output_graph=True # 是否输出 tensorboard 文件
)
env.after(100, run_maze)
env.mainloop()
RL.plot_cost() # 观看神经网络的误差曲线
4. Summary
Reinforcement learning itself does not depend on deep learning. It is more of an idea to update the Q table (or the neural network with the same function) by generating reward values through the interaction between behavior and the environment. It does not have a fixed code, only a set of patterns, and the specific code has to be written according to the actual application and interactive environment.