Hands-on ML and TF Chapter16 Reinforcement Learning

Policy Granients

import tensorflow as tf

reset_graph()

n_inputs = 4
n_hidden = 4
n_outputs = 1

learning_rate = 0.01

initializer = tf.variance_scaling_initializer()

X = tf.placeholder(tf.float32, shape=[None, n_inputs])

hidden = tf.layers.dense(X, n_hidden, activation=tf.nn.elu, kernel_initializer=initializer)
logits = tf.layers.dense(hidden, n_outputs)
outputs = tf.nn.sigmoid(logits)  # probability of action 0 (left)
p_left_and_right = tf.concat(axis=1, values=[outputs, 1 - outputs])
action = tf.multinomial(tf.log(p_left_and_right), num_samples=1)

y = 1. - tf.to_float(action)
cross_entropy = tf.nn.sigmoid_cross_entropy_with_logits(labels=y, logits=logits)
optimizer = tf.train.AdamOptimizer(learning_rate)
grads_and_vars = optimizer.compute_gradients(cross_entropy)
gradients = [grad for grad, variable in grads_and_vars]
gradient_placeholders = []
grads_and_vars_feed = []
for grad, variable in grads_and_vars:
    gradient_placeholder = tf.placeholder(tf.float32, shape=grad.get_shape())
    gradient_placeholders.append(gradient_placeholder)
    grads_and_vars_feed.append((gradient_placeholder, variable))
training_op = optimizer.apply_gradients(grads_and_vars_feed)

init = tf.global_variables_initializer()
saver = tf.train.Saver()

def discount_rewards(rewards, discount_rate):
    discounted_rewards = np.zeros(len(rewards))
    cumulative_rewards = 0
    for step in reversed(range(len(rewards))):
        cumulative_rewards = rewards[step] + cumulative_rewards * discount_rate
        discounted_rewards[step] = cumulative_rewards
    return discounted_rewards

def discount_and_normalize_rewards(all_rewards, discount_rate):
    all_discounted_rewards = [discount_rewards(rewards, discount_rate) for rewards in all_rewards]
    flat_rewards = np.concatenate(all_discounted_rewards)
    reward_mean = flat_rewards.mean()
    reward_std = flat_rewards.std()
    return [(discounted_rewards - reward_mean)/reward_std for discounted_rewards in all_discounted_rewards]

env = gym.make("CartPole-v0")

n_games_per_update = 10
n_max_steps = 1000
n_iterations = 250
save_iterations = 10
discount_rate = 0.95

with tf.Session() as sess:
    init.run()
    for iteration in range(n_iterations):
        print("\rIteration: {}".format(iteration), end="")
        all_rewards = []
        all_gradients = []
        for game in range(n_games_per_update):
            current_rewards = []
            current_gradients = []
            obs = env.reset()
            for step in range(n_max_steps):
                action_val, gradients_val = sess.run([action, gradients], feed_dict={X: obs.reshape(1, n_inputs)})
                obs, reward, done, info = env.step(action_val[0][0])
                current_rewards.append(reward)
                current_gradients.append(gradients_val)
                if done:
                    break
            all_rewards.append(current_rewards)
            all_gradients.append(current_gradients)

        all_rewards = discount_and_normalize_rewards(all_rewards, discount_rate=discount_rate)
        feed_dict = {}
        for var_index, gradient_placeholder in enumerate(gradient_placeholders):
            mean_gradients = np.mean([reward * all_gradients[game_index][step][var_index]
                                      for game_index, rewards in enumerate(all_rewards)
                                          for step, reward in enumerate(rewards)], axis=0)
            ## mean_gradients 是个标量，为一个Variable在10个episode的所有step的梯度的平均值。
            feed_dict[gradient_placeholder] = mean_gradients
        sess.run(training_op, feed_dict=feed_dict)
        if iteration % save_iterations == 0:
            saver.save(sess, "./my_policy_net_pg.ckpt")

Markov Decision Process

Learning to Play Ms. Pac-Man Using Deep Q-Learning

这部分首先解决在Windows下安装gym[atari]模块

查阅了很多资料，发现很多解决办法都是使用这个连接：

pip install --no-index -f https://github.com/Kojoley/atari-py/releases atari_py

这种的方法的出处是OpenAI Gym Atari on Windows

这个gym的github issue中也提到这个方法。

这个How to Install OpenAI Gym in a Windows Environment也是使用这个方法安装atari模块

也可以在下面这个网址上下载whl进行离线安装

https://github.com/Kojoley/atari-py/releases

对于安装whl格式的文件，首先要安装wheel包

pip install wheel

由于anconda中默认安装了wheel包，所以上面的步骤可以省略。

pip install atari_py-0.1.7-cp36-cp36m-win_amd64.whl

绘制智能体和环境交互的动画

# 导入需要的包
import matplotlib
import matplotlib.animation as animation
import matplotlib.pyplot as plt
plt.rcParams['axes.labelsize'] = 14
plt.rcParams['xtick.labelsize'] = 12
plt.rcParams['ytick.labelsize'] = 12

# 将每一帧的画面存到frames中
frames = []

n_max_steps = 1000
n_change_steps = 10

obs = env.reset()
for step in range(n_max_steps):
    img = env.render(mode="rgb_array")
    frames.append(img)
    if step % n_change_steps == 0:
        action = env.action_space.sample() # play randomly
    obs, reward, done, info = env.step(action)
    if done:
        break

# 定义绘制动画的函数
def update_scene(num, frames, patch):
    patch.set_data(frames[num])
    return patch,

def plot_animation(frames, repeat=False, interval=40):
    plt.close()  # or else nbagg sometimes plots in the previous cell
    fig = plt.figure()
    patch = plt.imshow(frames[0])
    plt.axis('off')
    return animation.FuncAnimation(fig, update_scene, fargs=(frames, patch), frames=len(frames), repeat=repeat, interval=interval)

# 调用函数显示动画
video = plot_animation(frames)
plt.show()

DQL算法部分

首先是对obs进行预处理，转换为88x80的灰度图像并增强对比度。

import gym
import matplotlib.pyplot as plt
import numpy as np

env = gym.make('MsPacman-v0')
obs = env.reset()
plt.imshow(obs)
plt.show()

mspacman_color = np.array([210, 164, 74]).mean()

def preprocess_observation(obs):
    img = obs[1:176:2,::2] # crop and downsize
    img = img.mean(axis=2) # to greyscale
    img[img == mspacman_color] = 0 # improve constrast
    img = (img - 128)/128 -1 # normalize from -1. to 1.
    return img.reshape(88,80,1)

pre_obs = preprocess_observation(obs).reshape(88,80)
plt.imshow(pre_obs, cmap ='gray')
plt.show()

DQN算法

# ------------------------------构建DQN的计算图-----------------------------------------
from tensorflow.contrib.layers import convolution2d, fully_connected

input_height = 88
input_width = 80
input_channels = 1
conv_n_maps = [32, 64, 64]
conv_kernel_sizes = [(8,8), (4,4), (3,3)]
conv_strides = [4, 2, 1]
conv_paddings = ["SAME"] * 3 
conv_activation = [tf.nn.relu] * 3
n_hidden_in = 64 * 11 * 10  # conv3 has 64 maps of 11x10 each
n_hidden = 512
hidden_activation = tf.nn.relu
n_outputs = env.action_space.n  # 9 discrete actions are available
initializer = tf.variance_scaling_initializer()

def q_network(X_state, name):
    prev_layer = X_state / 128.0 # scale pixel intensities to the [-1.0, 1.0] range.
    with tf.variable_scope(name) as scope:
        for n_maps, kernel_size, strides, padding, activation in zip(
                conv_n_maps, conv_kernel_sizes, conv_strides,
                conv_paddings, conv_activation):
            prev_layer = tf.layers.conv2d(
                prev_layer, filters=n_maps, kernel_size=kernel_size,
                strides=strides, padding=padding, activation=activation,
                kernel_initializer=initializer)
        last_conv_layer_flat = tf.reshape(prev_layer, shape=[-1, n_hidden_in])
        hidden = tf.layers.dense(last_conv_layer_flat, n_hidden,
                                 activation=hidden_activation,
                                 kernel_initializer=initializer)
        outputs = tf.layers.dense(hidden, n_outputs,
                                  kernel_initializer=initializer)
    trainable_vars = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES,
                                       scope=scope.name)
    trainable_vars_by_name = {var.name[len(scope.name):]: var
                              for var in trainable_vars}
    return outputs, trainable_vars_by_name

# --------------------------------两个DQN之间复制参数的过程----------------------------------
X_state = tf.placeholder(tf.float32, shape=[None, input_height, input_width,
                                            input_channels])
online_q_values, online_vars = q_network(X_state, name="q_networks/online")
target_q_values, target_vars = q_network(X_state, name="q_networks/target")

copy_ops = [target_var.assign(online_vars[var_name])
            for var_name, target_var in target_vars.items()]
copy_online_to_target = tf.group(*copy_ops)

# ------------------------------构建actor的损失函数和优化器
learning_rate = 0.001
momentum = 0.95

with tf.variable_scope("train"):
    X_action = tf.placeholder(tf.int32, shape=[None])
    y = tf.placeholder(tf.float32, shape=[None, 1])
    q_value = tf.reduce_sum(online_q_values * tf.one_hot(X_action, n_outputs),
                            axis=1, keepdims=True)
    error = tf.abs(y - q_value)
    clipped_error = tf.clip_by_value(error, 0.0, 1.0)
    linear_error = 2 * (error - clipped_error)
    loss = tf.reduce_mean(tf.square(clipped_error) + linear_error)

    global_step = tf.Variable(0, trainable=False, name='global_step')
    optimizer = tf.train.MomentumOptimizer(learning_rate, momentum, use_nesterov=True)
    training_op = optimizer.minimize(loss, global_step=global_step)

init = tf.global_variables_initializer()
saver = tf.train.Saver()

#----------------------------replay mem的设置代替书中过的deque-----------------------------------
class ReplayMemory:
    def __init__(self, maxlen):
        self.maxlen = maxlen
        self.buf = np.empty(shape=maxlen, dtype=np.object)
        self.index = 0
        self.length = 0
        
    def append(self, data):
        self.buf[self.index] = data
        self.length = min(self.length + 1, self.maxlen)
        self.index = (self.index + 1) % self.maxlen
    
    def sample(self, batch_size, with_replacement=True):
        if with_replacement:
            indices = np.random.randint(self.length, size=batch_size) # faster
        else:
            indices = np.random.permutation(self.length)[:batch_size]
        return self.buf[indices]
    
replay_memory_size = 500000
replay_memory = ReplayMemory(replay_memory_size)

#----------------------------------从replay mem中进行采样--------------------------------------
def sample_memories(batch_size):
    cols = [[], [], [], [], []] # state, action, reward, next_state, continue
    for memory in replay_memory.sample(batch_size):
        for col, value in zip(cols, memory):
            col.append(value)
    cols = [np.array(col) for col in cols]
    return cols[0], cols[1], cols[2].reshape(-1, 1), cols[3], cols[4].reshape(-1, 1)

#-------------------------------epsilon贪婪策略-----------------------------------------
eps_min = 0.1
eps_max = 1.0
eps_decay_steps = 2000000

def epsilon_greedy(q_values, step):
    epsilon = max(eps_min, eps_max - (eps_max-eps_min) * step/eps_decay_steps)
    if np.random.rand() < epsilon:
        return np.random.randint(n_outputs) # random action
    else:
        return np.argmax(q_values) # optimal action

#-------------------------------DQN的训练过程--------------------------------------------
n_steps = 4000000  # total number of training steps
training_start = 10000  # start training after 10,000 game iterations
training_interval = 4  # run a training step every 4 game iterations
save_steps = 1000  # save the model every 1,000 training steps
copy_steps = 10000  # copy online DQN to target DQN every 10,000 training steps
discount_rate = 0.99
skip_start = 90  # Skip the start of every game (it's just waiting time).
batch_size = 50
iteration = 0  # game iterations
checkpoint_path = "./my_dqn.ckpt"
done = True # env needs to be reset

loss_val = np.infty
game_length = 0
total_max_q = 0
mean_max_q = 0.0

with tf.Session() as sess:
    if os.path.isfile(checkpoint_path + ".index"):
        saver.restore(sess, checkpoint_path)
    else:
        init.run()
        copy_online_to_target.run()
    while True:
        step = global_step.eval()
        if step >= n_steps:
            break
        iteration += 1
        print("\rIteration {}\tTraining step {}/{} ({:.1f})%\tLoss {:5f}\tMean Max-Q {:5f}   ".format(
            iteration, step, n_steps, step * 100 / n_steps, loss_val, mean_max_q), end="")
        if done: # game over, start again
            obs = env.reset()
            for skip in range(skip_start): # skip the start of each game
                obs, reward, done, info = env.step(0)
            state = preprocess_observation(obs)

        # Online DQN evaluates what to do
        q_values = online_q_values.eval(feed_dict={X_state: [state]})
        action = epsilon_greedy(q_values, step)

        # Online DQN plays
        obs, reward, done, info = env.step(action)
        next_state = preprocess_observation(obs)

        # Let's memorize what happened
        replay_memory.append((state, action, reward, next_state, 1.0 - done))
        state = next_state

        # Compute statistics for tracking progress (not shown in the book)
        total_max_q += q_values.max()
        game_length += 1
        if done:
            mean_max_q = total_max_q / game_length
            total_max_q = 0.0
            game_length = 0

        if iteration < training_start or iteration % training_interval != 0:
            continue # only train after warmup period and at regular intervals
        
        # Sample memories and use the target DQN to produce the target Q-Value
        X_state_val, X_action_val, rewards, X_next_state_val, continues = (
            sample_memories(batch_size))
        next_q_values = target_q_values.eval(
            feed_dict={X_state: X_next_state_val})
        max_next_q_values = np.max(next_q_values, axis=1, keepdims=True)
        y_val = rewards + continues * discount_rate * max_next_q_values

        # Train the online DQN
        _, loss_val = sess.run([training_op, loss], feed_dict={
            X_state: X_state_val, X_action: X_action_val, y: y_val})

        # Regularly copy the online DQN to the target DQN
        if step % copy_steps == 0:
            copy_online_to_target.run()

        # And save regularly
        if step % save_steps == 0:
            saver.save(sess, checkpoint_path)