百度强化学习7日打卡营心得

最新推荐文章于 2020-07-24 19:36:50 发布

原创最新推荐文章于 2020-07-24 19:36:50 发布 · 581 阅读

0 ·

CC 4.0 BY-SA版权

文章标签：

#强化学习

RL 专栏收录该内容

1 篇文章

订阅专栏

本文详细介绍了参加百度强化学习7日打卡营的学习经历，涵盖Q-learning、DQN、PolicyGradient等经典算法的理论与实践，通过PARL框架实现并解决Pong游戏和四旋翼悬停任务。

最近参加了百度强化学习7日打卡营，从只是听说过Q-learning、AlphaGo的完全小白，到可以自己实现几个经典强化学习算法，并且使用百度PARL框架实现强化学习解决Pong游戏和四旋翼悬停。这次的7日打卡营让我快速入门强化学习，了解到几个经典强化学习算法并且动手实现，还是颇有收获的。在此也给百度PARL框架（Github）打个广告，PARL框架已经实现并封装好几个经典强化学习算法，如Q-learning、DQN、Policy Gradient、DDPG、PPO等，让开发者们可以专注于网络Model和Agent交互的设计，易用性和复用性非常高。（个人使用感觉使用PARL实现强化学习还是很方便的）此外，PARL还有大规模分布式能力，非常利于工业应用。（虽然没有用过这部分功能）

接下来就简单记录下这次课程的学习笔记。

Lesson 1

核心思想：智能体agent在环境environment中学习，根据环境的状态state（或观测到的observation），执行动作action，并根据环境的反馈 reward（奖励）来指导更好的动作。

经典算法：Q-learning、Sarsa、DQN、Policy Gradient、A3C、DDPG、PPO
环境分类：离散控制场景（输出动作可数）、连续控制场景（输出动作值不可数）
强化学习经典环境库GYM将环境交互接口规范化为：重置环境reset()、交互step()、渲染render()
强化学习框架库PARL将强化学习框架抽象为Model、Algorithm、Agent三层，使得强化学习算法的实现和调试更方便和灵活。

强化学习的一个关键是探索和利用的平衡。强化学习分类：基于价值value-based、基于策略policy-based。

Lesson 2

Sarsa

Sarsa全称是state-action-reward-state'-action'，目的是学习特定的state下，特定action的价值Q，最终建立和优化一个Q表格，以state为行，action为列，根据与环境交互得到的reward来更新Q表格，更新公式为：

$Q(S_t,A_t)\leftarrow Q(S_t,A_t)+\alpha[R_{t+1}+\gamma Q(S_{t+1},A_{t+1})-Q(S_t,A_t)]$

# agent.py 
class SarsaAgent(object):
    def __init__(self, obs_n, act_n, learning_rate=0.01, gamma=0.9, e_greed=0.1):
        self.act_n = act_n      # 动作维度，有几个动作可选
        self.lr = learning_rate # 学习率
        self.gamma = gamma      # reward的衰减率
        self.epsilon = e_greed  # 按一定概率随机选动作
        self.Q = np.zeros((obs_n, act_n))

    # 根据输入观察值，采样输出的动作值，带探索
    def sample(self, obs):
        if np.random.uniform(0, 1) < (1.0 - self.epsilon): #根据table的Q值选动作
            action = self.predict(obs)
        else:
            action = np.random.choice(self.act_n) #有一定概率随机探索选取一个动作
        return action

    # 根据输入观察值，预测输出的动作值
    def predict(self, obs):
        Q_list = self.Q[obs, :]
        maxQ = np.max(Q_list)
        action_list = np.where(Q_list == maxQ)[0]  # maxQ可能对应多个action
        action = np.random.choice(action_list)
        return action

    # 学习方法，也就是更新Q-table的方法
    def learn(self, obs, action, reward, next_obs, next_action, done):
        """ on-policy
            obs: 交互前的obs, s_t
            action: 本次交互选择的action, a_t
            reward: 本次动作获得的奖励r
            next_obs: 本次交互后的obs, s_t+1
            next_action: 根据当前Q表格, 针对next_obs会选择的动作, a_t+1
            done: episode是否结束
        """
        predict_Q = self.Q[obs, action]
        if done:
            target_Q = reward # 没有下一个状态了
        else:
            target_Q = reward + self.gamma * self.Q[next_obs, next_action] # Sarsa
        self.Q[obs, action] += self.lr * (target_Q - predict_Q) # 修正q

    # 保存Q表格数据到文件
    def save(self):
        npy_file = './q_table.npy'
        np.save(npy_file, self.Q)
        print(npy_file + ' saved.')
    
    # 从文件中读取Q值到Q表格中
    def restore(self, npy_file='./q_table.npy'):
        self.Q = np.load(npy_file)
        print(npy_file + ' loaded.')
        
        
def run_episode(env, agent, render=False):
    total_steps = 0 # 记录每个episode走了多少step
    total_reward = 0

    obs = env.reset() # 重置环境, 重新开一局（即开始新的一个episode）
    action = agent.sample(obs) # 根据算法选择一个动作

    while True:
        next_obs, reward, done, _ = env.step(action) # 与环境进行一个交互
        next_action = agent.sample(next_obs) # 根据算法选择一个动作
        # 训练 Sarsa 算法
        agent.learn(obs, action, reward, next_obs, next_action, done)

        action = next_action
        obs = next_obs  # 存储上一个观察值
        total_reward += reward
        total_steps += 1 # 计算step数
        if render:
            env.render() #渲染新的一帧图形
        if done:
            break
    return total_reward, total_steps


def test_episode(env, agent):
    total_reward = 0
    obs = env.reset()
    while True:
        action = agent.predict(obs) # greedy
        next_obs, reward, done, _ = env.step(action)
        total_reward += reward
        obs = next_obs
        # time.sleep(0.5)
        # env.render()
        if done:
            break
    return total_reward

Q-learning

Q-learning跟Sarsa不一样的地方是更新Q表格的方式。
- Sarsa是on-policy的更新方式，先做出动作再更新。
- Q-learning是off-policy的更新方式，更新learn()时无需获取下一步实际做出的动作next_action，并假设下一步动作是取最大Q值的动作。
Q-learning的更新公式为：
$Q(S_t,A_t)\leftarrow Q(S_t,A_t)+\alpha[R_{t+1}+\gamma \max_aQ(S_{t+1},a)-Q(S_t,A_t)]$

class QLearningAgent(object):
    def __init__(self, obs_n, act_n, learning_rate=0.01, gamma=0.9, e_greed=0.1):
        self.act_n = act_n      # 动作维度，有几个动作可选
        self.lr = learning_rate # 学习率
        self.gamma = gamma      # reward的衰减率
        self.epsilon = e_greed  # 按一定概率随机选动作
        self.Q = np.zeros((obs_n, act_n))

    # 根据输入观察值，采样输出的动作值，带探索
    def sample(self, obs):
        if np.random.uniform(0, 1) < (1.0 - self.epsilon): #根据table的Q值选动作
            action = self.predict(obs)
        else:
            action = np.random.choice(self.act_n) #有一定概率随机探索选取一个动作
        return action

    # 根据输入观察值，预测输出的动作值
    def predict(self, obs):
        Q_list = self.Q[obs, :]
        maxQ = np.max(Q_list)
        action_list = np.where(Q_list == maxQ)[0]  # maxQ可能对应多个action
        action = np.random.choice(action_list)
        return action

    # 学习方法，也就是更新Q-table的方法
    def learn(self, obs, action, reward, next_obs, done):
        """ off-policy
            obs: 交互前的obs, s_t
            action: 本次交互选择的action, a_t
            reward: 本次动作获得的奖励r
            next_obs: 本次交互后的obs, s_t+1
            done: episode是否结束
        """
        predict_Q = self.Q[obs, action]
        if done:
            target_Q = reward # 没有下一个状态了
        else:
            target_Q = reward + self.gamma * np.max(self.Q[next_obs, :]) # Q-learning
        self.Q[obs, action] += self.lr * (target_Q - predict_Q) # 修正q

    # 把 Q表格 的数据保存到文件中
    def save(self):
        npy_file = './q_table.npy'
        np.save(npy_file, self.Q)
        print(npy_file + ' saved.')
    
    # 从文件中读取数据到 Q表格
    def restore(self, npy_file='./q_table.npy'):
        self.Q = np.load(npy_file)
        print(npy_file + ' loaded.')
        
      
def run_episode(env, agent, render=False):
    total_steps = 0 # 记录每个episode走了多少step
    total_reward = 0

    obs = env.reset() # 重置环境, 重新开一局（即开始新的一个episode）

    while True:
        action = agent.sample(obs) # 根据算法选择一个动作
        next_obs, reward, done, _ = env.step(action) # 与环境进行一个交互
        # 训练 Q-learning算法
        agent.learn(obs, action, reward, next_obs, done)

        obs = next_obs  # 存储上一个观察值
        total_reward += reward
        total_steps += 1 # 计算step数
        if render:
            env.render() #渲染新的一帧图形
        if done:
            break
    return total_reward, total_steps

def test_episode(env, agent):
    total_reward = 0
    obs = env.reset()
    while True:
        action = agent.predict(obs) # greedy
        next_obs, reward, done, _ = env.step(action)
        total_reward += reward
        obs = next_obs
        # time.sleep(0.5)
        # env.render()
        if done:
            break
    return total_reward

On-policy策略：使用策略π学习，使用策略π与环境交互产生经验。由于需要兼顾探索，策略π并不稳定

Off-policy策略：目标策略π；行为策略μ。目标策略用于学习最优策略，行为策略更具有探索性，与环境交互产生经验轨迹

Lesson 3

DQN

本质上DQN还是一个Q-learning算法，更新方式一致。为了更好的探索环境，同样的也采用ε-greedy方法训练。
在Q-learning的基础上，DQN提出了两个技巧使得Q网络的更新迭代更稳定。
- 经验回放 Experience Replay：主要解决样本关联性和利用效率的问题。使用一个经验池存储多条经验s,a,r,s'，再从中随机抽取一批数据送去训练。
- 固定Q目标 Fixed-Q-Target：主要解决算法训练不稳定的问题。复制一个和原来Q网络结构一样的Target Q网络，用于计算Q目标值。

class Model(parl.Model):
    def __init__(self, act_dim):
        hid1_size = 128
        hid2_size = 128
        # 3层全连接网络
        self.fc1 = layers.fc(size=hid1_size, act='relu')
        self.fc2 = layers.fc(size=hid2_size, act='relu')
        self.fc3 = layers.fc(size=act_dim, act=None)

    def value(self, obs):
        # 定义网络
        # 输入state，输出所有action对应的Q，[Q(s,a1), Q(s,a2), Q(s,a3)...]
        h1 = self.fc1(obs)
        h2 = self.fc2(h1)
        Q = self.fc3(h2)
        return Q
    
class DQN(parl.Algorithm):
    def __init__(self, model, act_dim=None, gamma=None, lr=None):
        """ DQN algorithm
        
        Args:
            model (parl.Model): 定义Q函数的前向网络结构
            act_dim (int): action空间的维度，即有几个action
            gamma (float): reward的衰减因子
            lr (float): learning rate 学习率.
        """
        self.model = model
        self.target_model = copy.deepcopy(model)

        assert isinstance(act_dim, int)
        assert isinstance(gamma, float)
        assert isinstance(lr, float)
        self.act_dim = act_dim
        self.gamma = gamma
        self.lr = lr

    def predict(self, obs):
        """ 使用self.model的value网络来获取 [Q(s,a1),Q(s,a2),...]
        """
        return self.model.value(obs)

    def learn(self, obs, action, reward, next_obs, terminal):
        """ 使用DQN算法更新self.model的value网络
        """
        # 从target_model中获取 max Q' 的值，用于计算target_Q
        next_pred_value = self.target_model.value(next_obs)
        best_v = layers.reduce_max(next_pred_value, dim=1)
        best_v.stop_gradient = True  # 阻止梯度传递
        terminal = layers.cast(terminal, dtype='float32')
        target = reward + (1.0 - terminal) * self.gamma * best_v

        pred_value = self.model.value(obs)  # 获取Q预测值
        # 将action转onehot向量，比如：3 => [0,0,0,1,0]
        action_onehot = layers.one_hot(action, self.act_dim)
        action_onehot = layers.cast(action_onehot, dtype='float32')
        # 下面一行是逐元素相乘，拿到action对应的 Q(s,a)
        # 比如：pred_value = [[2.3, 5.7, 1.2, 3.9, 1.4]], action_onehot = [[0,0,0,1,0]]
        #  ==> pred_action_value = [[3.9]]
        pred_action_value = layers.reduce_sum(
            layers.elementwise_mul(action_onehot, pred_value), dim=1)

        # 计算 Q(s,a) 与 target_Q的均方差，得到loss
        cost = layers.square_error_cost(pred_action_value, target)
        cost = layers.reduce_mean(cost)
        optimizer = fluid.optimizer.Adam(learning_rate=self.lr)  # 使用Adam优化器
        optimizer.minimize(cost)
        return cost

    def sync_target(self):
        """ 把 self.model 的模型参数值同步到 self.target_model
        """
        self.model.sync_weights_to(self.target_model)
        
class Agent(parl.Agent):
    def __init__(self,
                 algorithm,
                 obs_dim,
                 act_dim,
                 e_greed=0.1,
                 e_greed_decrement=0):
        assert isinstance(obs_dim, int)
        assert isinstance(act_dim, int)
        self.obs_dim = obs_dim
        self.act_dim = act_dim
        super(Agent, self).__init__(algorithm)

        self.global_step = 0
        self.update_target_steps = 200  # 每隔200个training steps再把model的参数复制到target_model中

        self.e_greed = e_greed  # 有一定概率随机选取动作，探索
        self.e_greed_decrement = e_greed_decrement  # 随着训练逐步收敛，探索的程度慢慢降低

    def build_program(self):
        self.pred_program = fluid.Program()
        self.learn_program = fluid.Program()

        with fluid.program_guard(self.pred_program):  # 搭建计算图用于 预测动作，定义输入输出变量
            obs = layers.data(
                name='obs', shape=[self.obs_dim], dtype='float32')
            self.value = self.alg.predict(obs)

        with fluid.program_guard(self.learn_program):  # 搭建计算图用于 更新Q网络，定义输入输出变量
            obs = layers.data(
                name='obs', shape=[self.obs_dim], dtype='float32')
            action = layers.data(name='act', shape=[1], dtype='int32')
            reward = layers.data(name='reward', shape=[], dtype='float32')
            next_obs = layers.data(
                name='next_obs', shape=[self.obs_dim], dtype='float32')
            terminal = layers.data(name='terminal', shape=[], dtype='bool')
            self.cost = self.alg.learn(obs, action, reward, next_obs, terminal)

    def sample(self, obs):
        sample = np.random.rand()  # 产生0~1之间的小数
        if sample < self.e_greed:
            act = np.random.randint(self.act_dim)  # 探索：每个动作都有概率被选择
        else:
            act = self.predict(obs)  # 选择最优动作
        self.e_greed = max(
            0.01, self.e_greed - self.e_greed_decrement)  # 随着训练逐步收敛，探索的程度慢慢降低
        return act

    def predict(self, obs):  # 选择最优动作
        obs = np.expand_dims(obs, axis=0)
        pred_Q = self.fluid_executor.run(
            self.pred_program,
            feed={'obs': obs.astype('float32')},
            fetch_list=[self.value])[0]
        pred_Q = np.squeeze(pred_Q, axis=0)
        act = np.argmax(pred_Q)  # 选择Q最大的下标，即对应的动作
        return act

    def learn(self, obs, act, reward, next_obs, terminal):
        # 每隔200个training steps同步一次model和target_model的参数
        if self.global_step % self.update_target_steps == 0:
            self.alg.sync_target()
        self.global_step += 1

        act = np.expand_dims(act, -1)
        feed = {
            'obs': obs.astype('float32'),
            'act': act.astype('int32'),
            'reward': reward,
            'next_obs': next_obs.astype('float32'),
            'terminal': terminal
        }
        cost = self.fluid_executor.run(
            self.learn_program, feed=feed, fetch_list=[self.cost])[0]  # 训练一次网络
        return cost
    
import random
import collections
import numpy as np


class ReplayMemory(object):
    def __init__(self, max_size):
        self.buffer = collections.deque(maxlen=max_size)

    # 增加一条经验到经验池中
    def append(self, exp):
        self.buffer.append(exp)

    # 从经验池中选取N条经验出来
    def sample(self, batch_size):
        mini_batch = random.sample(self.buffer, batch_size)
        obs_batch, action_batch, reward_batch, next_obs_batch, done_batch = [], [], [], [], []

        for experience in mini_batch:
            s, a, r, s_p, done = experience
            obs_batch.append(s)
            action_batch.append(a)
            reward_batch.append(r)
            next_obs_batch.append(s_p)
            done_batch.append(done)

        return np.array(obs_batch).astype('float32'), \
            np.array(action_batch).astype('float32'), np.array(reward_batch).astype('float32'),\
            np.array(next_obs_batch).astype('float32'), np.array(done_batch).astype('float32')

    def __len__(self):
        return len(self.buffer)
    
# 训练一个episode
def run_episode(env, agent, rpm):
    total_reward = 0
    obs = env.reset()
    step = 0
    while True:
        step += 1
        action = agent.sample(obs)  # 采样动作，所有动作都有概率被尝试到
        next_obs, reward, done, _ = env.step(action)
        rpm.append((obs, action, reward, next_obs, done))

        # train model
        if (len(rpm) > MEMORY_WARMUP_SIZE) and (step % LEARN_FREQ == 0):
            (batch_obs, batch_action, batch_reward, batch_next_obs,
             batch_done) = rpm.sample(BATCH_SIZE)
            train_loss = agent.learn(batch_obs, batch_action, batch_reward,
                                     batch_next_obs,
                                     batch_done)  # s,a,r,s',done

        total_reward += reward
        obs = next_obs
        if done:
            break
    return total_reward


# 评估 agent, 跑 5 个episode，总reward求平均
def evaluate(env, agent, render=False):
    eval_reward = []
    for i in range(5):
        obs = env.reset()
        episode_reward = 0
        while True:
            action = agent.predict(obs)  # 预测动作，只选最优动作
            obs, reward, done, _ = env.step(action)
            episode_reward += reward
            if render:
                env.render()
            if done:
                break
        eval_reward.append(episode_reward)
    return np.mean(eval_reward)

Lesson 4

Policy Gradient

采用神经网络拟合策略函数，需计算策略梯度用于优化策略网络。

优化的目标是在策略π(s,a)的期望回报：所有的轨迹获得的回报R与对应的轨迹发生概率p的加权和，当N足够大时，可通过采样N个Episode求平均的方式近似表达。

$\bar R_\theta=\sum_\tau R(\tau)p_\theta(\tau)\approx\frac 1N\sum R(\tau)$

优化目标对参数θ求导后得到策略梯度：

$\nabla\bar R_\theta\approx\frac 1N\sum_{n}\sum_{t}R(\tau^n)\nabla\log\pi_\theta(a_t^n|s_t^n)$

class Model(parl.Model):
    def __init__(self, act_dim):
        act_dim = act_dim
        hid1_size = act_dim * 10

        self.fc1 = layers.fc(size=hid1_size, act='tanh')
        self.fc2 = layers.fc(size=act_dim, act='softmax')

    def forward(self, obs):  # 可直接用 model = Model(5); model(obs)调用
        out = self.fc1(obs)
        out = self.fc2(out)
        return out
    
# from parl.algorithms import PolicyGradient # 也可以直接从parl库中导入PolicyGradient算法，无需重复写算法

class PolicyGradient(parl.Algorithm):
    def __init__(self, model, lr=None):
        """ Policy Gradient algorithm
        
        Args:
            model (parl.Model): policy的前向网络.
            lr (float): 学习率.
        """

        self.model = model
        assert isinstance(lr, float)
        self.lr = lr

    def predict(self, obs):
        """ 使用policy model预测输出的动作概率
        """
        return self.model(obs)

    def learn(self, obs, action, reward):
        """ 用policy gradient 算法更新policy model
        """
        act_prob = self.model(obs)  # 获取输出动作概率
        # log_prob = layers.cross_entropy(act_prob, action) # 交叉熵
        log_prob = layers.reduce_sum(
            -1.0 * layers.log(act_prob) * layers.one_hot(
                action, act_prob.shape[1]),
            dim=1)
        cost = log_prob * reward
        cost = layers.reduce_mean(cost)

        optimizer = fluid.optimizer.Adam(self.lr)
        optimizer.minimize(cost)
        return cost

class Agent(parl.Agent):
    def __init__(self, algorithm, obs_dim, act_dim):
        self.obs_dim = obs_dim
        self.act_dim = act_dim
        super(Agent, self).__init__(algorithm)

    def build_program(self):
        self.pred_program = fluid.Program()
        self.learn_program = fluid.Program()

        with fluid.program_guard(self.pred_program):  # 搭建计算图用于 预测动作，定义输入输出变量
            obs = layers.data(
                name='obs', shape=[self.obs_dim], dtype='float32')
            self.act_prob = self.alg.predict(obs)

        with fluid.program_guard(
                self.learn_program):  # 搭建计算图用于 更新policy网络，定义输入输出变量
            obs = layers.data(
                name='obs', shape=[self.obs_dim], dtype='float32')
            act = layers.data(name='act', shape=[1], dtype='int64')
            reward = layers.data(name='reward', shape=[], dtype='float32')
            self.cost = self.alg.learn(obs, act, reward)

    def sample(self, obs):
        obs = np.expand_dims(obs, axis=0)  # 增加一维维度
        act_prob = self.fluid_executor.run(
            self.pred_program,
            feed={'obs': obs.astype('float32')},
            fetch_list=[self.act_prob])[0]
        act_prob = np.squeeze(act_prob, axis=0)  # 减少一维维度
        act = np.random.choice(range(self.act_dim), p=act_prob)  # 根据动作概率选取动作
        return act

    def predict(self, obs):
        obs = np.expand_dims(obs, axis=0)
        act_prob = self.fluid_executor.run(
            self.pred_program,
            feed={'obs': obs.astype('float32')},
            fetch_list=[self.act_prob])[0]
        act_prob = np.squeeze(act_prob, axis=0)
        act = np.argmax(act_prob)  # 根据动作概率选择概率最高的动作
        return act

    def learn(self, obs, act, reward):
        act = np.expand_dims(act, axis=-1)
        feed = {
            'obs': obs.astype('float32'),
            'act': act.astype('int64'),
            'reward': reward.astype('float32')
        }
        cost = self.fluid_executor.run(
            self.learn_program, feed=feed, fetch_list=[self.cost])[0]
        return cost

def run_episode(env, agent):
    obs_list, action_list, reward_list = [], [], []
    obs = env.reset()
    while True:
        obs_list.append(obs)
        action = agent.sample(obs) # 采样动作
        action_list.append(action)

        obs, reward, done, info = env.step(action)
        reward_list.append(reward)

        if done:
            break
    return obs_list, action_list, reward_list

# 评估 agent, 跑 5 个episode，总reward求平均
def evaluate(env, agent, render=False):
    eval_reward = []
    for i in range(5):
        obs = env.reset()
        episode_reward = 0
        while True:
            action = agent.predict(obs) # 选取最优动作
            obs, reward, isOver, _ = env.step(action)
            episode_reward += reward
            if render:
                env.render()
            if isOver:
                break
        eval_reward.append(episode_reward)
    return np.mean(eval_reward)

# 根据一个episode的每个step的reward列表，计算每一个Step的Gt
def calc_reward_to_go(reward_list, gamma=1.0):
    for i in range(len(reward_list) - 2, -1, -1):
        # G_t = r_t + γ·r_t+1 + ... = r_t + γ·G_t+1
        reward_list[i] += gamma * reward_list[i + 1]  # Gt
    return np.array(reward_list)


# 创建环境
env = gym.make('CartPole-v0')
obs_dim = env.observation_space.shape[0]
act_dim = env.action_space.n
logger.info('obs_dim {}, act_dim {}'.format(obs_dim, act_dim))

# 根据parl框架构建agent
model = Model(act_dim=act_dim)
alg = PolicyGradient(model, lr=LEARNING_RATE)
agent = Agent(alg, obs_dim=obs_dim, act_dim=act_dim)

# 加载模型
# if os.path.exists('./model.ckpt'):
#     agent.restore('./model.ckpt')
#     run_episode(env, agent, train_or_test='test', render=True)
#     exit()

for i in range(1000):
    obs_list, action_list, reward_list = run_episode(env, agent)
    if i % 10 == 0:
        logger.info("Episode {}, Reward Sum {}.".format(
            i, sum(reward_list)))

    batch_obs = np.array(obs_list)
    batch_action = np.array(action_list)
    batch_reward = calc_reward_to_go(reward_list)

    agent.learn(batch_obs, batch_action, batch_reward)
    if (i + 1) % 100 == 0:
        total_reward = evaluate(env, agent, render=False) # render=True 查看渲染效果，需要在本地运行，AIStudio无法显示
        logger.info('Test reward: {}'.format(total_reward))

# 保存模型到文件 ./model.ckpt
agent.save('./model.ckpt')

Lesson 5

DDPG(Deep Deterministic Policy Gradient)

DDPG的提出动机其实是为了让DQN可以扩展到连续的动作空间。DDPG借鉴了DQN的两个技巧：经验回放和固定Q网络。DDPG使用策略网络直接输出确定性动作。DDPG使用了单步更新的Actor-Critic的架构。

Target Network参数软更新 $\bar\theta\leftarrow\tau\theta+(1-\tau)\bar\theta$ ，使target network更稳定不至于变化太大。

ACTOR_LR = 1e-3  # Actor网络的 learning rate
CRITIC_LR = 1e-3  # Critic网络的 learning rate

GAMMA = 0.99      # reward 的衰减因子
TAU = 0.001       # 软更新的系数
MEMORY_SIZE = int(1e6)                  # 经验池大小
MEMORY_WARMUP_SIZE = MEMORY_SIZE // 20  # 预存一部分经验之后再开始训练
BATCH_SIZE = 128
REWARD_SCALE = 0.1   # reward 缩放系数
NOISE = 0.05         # 动作噪声方差

TRAIN_EPISODE = 6000 # 训练的总episode数

class Model(parl.Model):
    def __init__(self, act_dim):
        self.actor_model = ActorModel(act_dim)
        self.critic_model = CriticModel()

    def policy(self, obs):
        return self.actor_model.policy(obs)

    def value(self, obs, act):
        return self.critic_model.value(obs, act)

    def get_actor_params(self):
        return self.actor_model.parameters()


class ActorModel(parl.Model):
    def __init__(self, act_dim):
        hid_size = 100

        self.fc1 = layers.fc(size=hid_size, act='relu')
        self.fc2 = layers.fc(size=act_dim, act='tanh')

    def policy(self, obs):
        hid = self.fc1(obs)
        means = self.fc2(hid)
        return means


class CriticModel(parl.Model):
    def __init__(self):
        hid_size = 100

        self.fc1 = layers.fc(size=hid_size, act='relu')
        self.fc2 = layers.fc(size=1, act=None)

    def value(self, obs, act):
        concat = layers.concat([obs, act], axis=1)
        hid = self.fc1(concat)
        Q = self.fc2(hid)
        Q = layers.squeeze(Q, axes=[1])
        return Q
    
class DDPG(parl.Algorithm):
    def __init__(self,
                 model,
                 gamma=None,
                 tau=None,
                 actor_lr=None,
                 critic_lr=None):
        """  DDPG algorithm
        
        Args:
            model (parl.Model): actor and critic 的前向网络.
                                model 必须实现 get_actor_params() 方法.
            gamma (float): reward的衰减因子.
            tau (float): self.target_model 跟 self.model 同步参数 的 软更新参数
            actor_lr (float): actor 的学习率
            critic_lr (float): critic 的学习率
        """
        assert isinstance(gamma, float)
        assert isinstance(tau, float)
        assert isinstance(actor_lr, float)
        assert isinstance(critic_lr, float)
        self.gamma = gamma
        self.tau = tau
        self.actor_lr = actor_lr
        self.critic_lr = critic_lr

        self.model = model
        self.target_model = deepcopy(model)

    def predict(self, obs):
        """ 使用 self.model 的 actor model 来预测动作
        """
        return self.model.policy(obs)

    def learn(self, obs, action, reward, next_obs, terminal):
        """ 用DDPG算法更新 actor 和 critic
        """
        actor_cost = self._actor_learn(obs)
        critic_cost = self._critic_learn(obs, action, reward, next_obs,
                                         terminal)
        return actor_cost, critic_cost

    def _actor_learn(self, obs):
        action = self.model.policy(obs)
        Q = self.model.value(obs, action)
        cost = layers.reduce_mean(-1.0 * Q)
        optimizer = fluid.optimizer.AdamOptimizer(self.actor_lr)
        optimizer.minimize(cost, parameter_list=self.model.get_actor_params())
        return cost

    def _critic_learn(self, obs, action, reward, next_obs, terminal):
        next_action = self.target_model.policy(next_obs)
        next_Q = self.target_model.value(next_obs, next_action)

        terminal = layers.cast(terminal, dtype='float32')
        target_Q = reward + (1.0 - terminal) * self.gamma * next_Q
        target_Q.stop_gradient = True

        Q = self.model.value(obs, action)
        cost = layers.square_error_cost(Q, target_Q)
        cost = layers.reduce_mean(cost)
        optimizer = fluid.optimizer.AdamOptimizer(self.critic_lr)
        optimizer.minimize(cost)
        return cost

    def sync_target(self, decay=None, share_vars_parallel_executor=None):
        """ self.target_model从self.model复制参数过来，可设置软更新参数
        """
        if decay is None:
            decay = 1.0 - self.tau
        self.model.sync_weights_to(
            self.target_model,
            decay=decay,
            share_vars_parallel_executor=share_vars_parallel_executor)
        
class Agent(parl.Agent):
    def __init__(self, algorithm, obs_dim, act_dim):
        assert isinstance(obs_dim, int)
        assert isinstance(act_dim, int)
        self.obs_dim = obs_dim
        self.act_dim = act_dim
        super(Agent, self).__init__(algorithm)

        # 注意：最开始先同步self.model和self.target_model的参数.
        self.alg.sync_target(decay=0)

    def build_program(self):
        self.pred_program = fluid.Program()
        self.learn_program = fluid.Program()

        with fluid.program_guard(self.pred_program):
            obs = layers.data(
                name='obs', shape=[self.obs_dim], dtype='float32')
            self.pred_act = self.alg.predict(obs)

        with fluid.program_guard(self.learn_program):
            obs = layers.data(
                name='obs', shape=[self.obs_dim], dtype='float32')
            act = layers.data(
                name='act', shape=[self.act_dim], dtype='float32')
            reward = layers.data(name='reward', shape=[], dtype='float32')
            next_obs = layers.data(
                name='next_obs', shape=[self.obs_dim], dtype='float32')
            terminal = layers.data(name='terminal', shape=[], dtype='bool')
            _, self.critic_cost = self.alg.learn(obs, act, reward, next_obs,
                                                 terminal)

    def predict(self, obs):
        obs = np.expand_dims(obs, axis=0)
        act = self.fluid_executor.run(
            self.pred_program, feed={'obs': obs},
            fetch_list=[self.pred_act])[0]
        act = np.squeeze(act)
        return act

    def learn(self, obs, act, reward, next_obs, terminal):
        feed = {
            'obs': obs,
            'act': act,
            'reward': reward,
            'next_obs': next_obs,
            'terminal': terminal
        }
        critic_cost = self.fluid_executor.run(
            self.learn_program, feed=feed, fetch_list=[self.critic_cost])[0]
        self.alg.sync_target()
        return critic_cost
    
def run_episode(agent, env, rpm):
    obs = env.reset()
    total_reward = 0
    steps = 0
    while True:
        steps += 1
        batch_obs = np.expand_dims(obs, axis=0)
        action = agent.predict(batch_obs.astype('float32'))

        # 增加探索扰动, 输出限制在 [-1.0, 1.0] 范围内
        action = np.clip(np.random.normal(action, NOISE), -1.0, 1.0)

        next_obs, reward, done, info = env.step(action)

        action = [action]  # 方便存入replaymemory
        rpm.append((obs, action, REWARD_SCALE * reward, next_obs, done))

        if len(rpm) > MEMORY_WARMUP_SIZE and (steps % 5) == 0:
            (batch_obs, batch_action, batch_reward, batch_next_obs,
             batch_done) = rpm.sample(BATCH_SIZE)
            agent.learn(batch_obs, batch_action, batch_reward, batch_next_obs,
                        batch_done)

        obs = next_obs
        total_reward += reward

        if done or steps >= 200:
            break
    return total_reward

import random
import collections
import numpy as np


class ReplayMemory(object):
    def __init__(self, max_size):
        self.buffer = collections.deque(maxlen=max_size)

    def append(self, exp):
        self.buffer.append(exp)

    def sample(self, batch_size):
        mini_batch = random.sample(self.buffer, batch_size)
        obs_batch, action_batch, reward_batch, next_obs_batch, done_batch = [], [], [], [], []

        for experience in mini_batch:
            s, a, r, s_p, done = experience
            obs_batch.append(s)
            action_batch.append(a)
            reward_batch.append(r)
            next_obs_batch.append(s_p)
            done_batch.append(done)

        return np.array(obs_batch).astype('float32'), \
            np.array(action_batch).astype('float32'), np.array(reward_batch).astype('float32'),\
            np.array(next_obs_batch).astype('float32'), np.array(done_batch).astype('float32')

    def __len__(self):
        return len(self.buffer)

def evaluate(env, agent, render=False):
    eval_reward = []
    for i in range(5):
        obs = env.reset()
        total_reward = 0
        steps = 0
        while True:
            batch_obs = np.expand_dims(obs, axis=0)
            action = agent.predict(batch_obs.astype('float32'))
            action = np.clip(action, -1.0, 1.0)

            steps += 1
            next_obs, reward, done, info = env.step(action)

            obs = next_obs
            total_reward += reward

            if render:
                env.render()
            if done or steps >= 200:
                break
        eval_reward.append(total_reward)
    return np.mean(eval_reward)

总结：学习强化学习算法还是需要结合理论和代码才能更深入理解算法细节，更重要的是实践，个人感觉强化学习的模型设计以及训练调参还是比较玄学的，不像监督学习训练时可以通过学习曲线预见模型的收敛性，强化学习模型训练时间似乎也更长，利用经验和直觉设计好的模型和适合的超参会使得强化学习顺利很多（在作业实战中我就是设计了不怎么好的模型以及超参，导致训练了好久也没收敛到理想值，不得已又要重头开始，花费了相当多的时间）。