深度 Q 网络：DQN 项目实战 CartPole-v0

2023-06-25
广东
本文字数：6471 字
阅读完需：约 21 分钟

本文分享自华为云社区《强化学习从基础到进阶-案例与实践[4.1]：深度Q网络-DQN项目实战CartPole-v0》，作者：汀丶。

1、定义算法

相比于 Q learning，DQN 本质上是为了适应更为复杂的环境，并且经过不断的改良迭代，到了 Nature DQN（即 Volodymyr Mnih 发表的 Nature 论文）这里才算是基本完善。DQN 主要改动的点有三个：

使用深度神经网络替代原来的 Q 表：这个很容易理解原因
使用了经验回放（Replay Buffer）：这个好处有很多，一个是使用一堆历史数据去训练，比之前用一次就扔掉好多了，大大提高样本效率，另外一个是面试常提到的，减少样本之间的相关性，原则上获取经验跟学习阶段是分开的，原来时序的训练数据有可能是不稳定的，打乱之后再学习有助于提高训练的稳定性，跟深度学习中划分训练测试集时打乱样本是一个道理。
使用了两个网络：即策略网络和目标网络，每隔若干步才把每步更新的策略网络参数复制给目标网络，这样做也是为了训练的稳定，避免 Q 值的估计发散。想象一下，如果当前有个 transition（这个 Q learning 中提过的，一定要记住！！！）样本导致对 Q 值进行了较差的过估计，如果接下来从经验回放中提取到的样本正好连续几个都这样的，很有可能导致 Q 值的发散（它的青春小鸟一去不回来了）。再打个比方，我们玩 RPG 或者闯关类游戏，有些人为了破纪录经常 Save 和 Load，只要我出了错，我不满意我就加载之前的存档，假设不允许加载呢，就像 DQN 算法一样训练过程中会退不了，这时候是不是搞两个档，一个档每帧都存一下，另外一个档打了不错的结果再存，也就是若干个间隔再存一下，到最后用间隔若干步数再存的档一般都比每帧都存的档好些呢。当然你也可以再搞更多个档，也就是 DQN 增加多个目标网络，但是对于 DQN 则没有多大必要，多几个网络效果不见得会好很多。

1.1 定义模型

import paddleimport paddle.nn as nnimport paddle.nn.functional as F!pip uninstall -y parl!pip install parlimport parlfrom parl.algorithms import DQNclass MLP(parl.Model): """ Linear network to solve Cartpole problem. Args: input_dim (int): Dimension of observation space. output_dim (int): Dimension of action space.    """ def __init__(self, input_dim, output_dim): super(MLP, self).__init__()        hidden_dim1 = 256        hidden_dim2 = 256        self.fc1 = nn.Linear(input_dim, hidden_dim1)        self.fc2 = nn.Linear(hidden_dim1, hidden_dim2)        self.fc3 = nn.Linear(hidden_dim2, output_dim) def forward(self, state):        x = F.relu(self.fc1(state))        x = F.relu(self.fc2(x))        x = self.fc3(x) return x

复制代码

1.2 定义经验回放

from collections import dequeclass ReplayBuffer: def __init__(self, capacity: int) -> None: self.capacity = capacity self.buffer = deque(maxlen=self.capacity) def push(self,transitions): '''_summary_ Args: trainsitions (tuple): _description_        ''' self.buffer.append(transitions) def sample(self, batch_size: int, sequential: bool = False): if batch_size > len(self.buffer): batch_size = len(self.buffer) if sequential: # sequential sampling            rand = random.randint(0, len(self.buffer) - batch_size)            batch = [self.buffer[i] for i in range(rand, rand + batch_size)] return zip(*batch) else:            batch = random.sample(self.buffer, batch_size) return zip(*batch) def clear(self): self.buffer.clear() def __len__(self): return len(self.buffer)

复制代码

1.3 定义智能体

from random import randomimport parlimport paddleimport mathimport numpy as npclass DQNAgent(parl.Agent): """Agent of DQN.    """ def __init__(self, algorithm, memory,cfg): super(DQNAgent, self).__init__(algorithm) self.n_actions = cfg['n_actions'] self.epsilon = cfg['epsilon_start'] self.sample_count = 0  self.epsilon_start = cfg['epsilon_start'] self.epsilon_end = cfg['epsilon_end'] self.epsilon_decay = cfg['epsilon_decay'] self.batch_size = cfg['batch_size'] self.global_step = 0 self.update_target_steps = 600 self.memory = memory # replay buffer def sample_action(self, state): self.sample_count += 1 # epsilon must decay(linear,exponential and etc.) for balancing exploration and exploitation self.epsilon = self.epsilon_end + (self.epsilon_start - self.epsilon_end) * \ math.exp(-1. * self.sample_count / self.epsilon_decay)  if random.random() < self.epsilon:            action = np.random.randint(self.n_actions) else:            action = self.predict_action(state) return action def predict_action(self, state):        state = paddle.to_tensor(state , dtype='float32') q_values = self.alg.predict(state) # self.alg 是自带的算法        action = q_values.argmax().numpy()[0] return action def update(self): """Update model with an episode data Args: obs(np.float32): shape of (batch_size, obs_dim)            act(np.int32): shape of (batch_size)            reward(np.float32): shape of (batch_size) next_obs(np.float32): shape of (batch_size, obs_dim)            terminal(np.float32): shape of (batch_size)        Returns:            loss(float)        """ if len(self.memory) < self.batch_size: # when transitions in memory donot meet a batch, not update return if self.global_step % self.update_target_steps == 0: self.alg.sync_target() self.global_step += 1 state_batch, action_batch, reward_batch, next_state_batch, done_batch = self.memory.sample( self.batch_size) action_batch = np.expand_dims(action_batch, axis=-1) reward_batch = np.expand_dims(reward_batch, axis=-1) done_batch = np.expand_dims(done_batch, axis=-1) state_batch = paddle.to_tensor(state_batch, dtype='float32') action_batch = paddle.to_tensor(action_batch, dtype='int32') reward_batch = paddle.to_tensor(reward_batch, dtype='float32') next_state_batch = paddle.to_tensor(next_state_batch, dtype='float32') done_batch = paddle.to_tensor(done_batch, dtype='float32')        loss = self.alg.learn(state_batch, action_batch, reward_batch, next_state_batch, done_batch)

复制代码

2、定义训练

def train(cfg, env, agent): ''' 训练    ''' print(f"开始训练！") print(f"环境：{cfg['env_name']}，算法：{cfg['algo_name']}，设备：{cfg['device']}")    rewards = [] # record rewards for all episodes    steps = [] for i_ep in range(cfg["train_eps"]): ep_reward = 0 # reward per episode ep_step = 0        state = env.reset() # reset and obtain initial state for _ in range(cfg['ep_max_steps']): ep_step += 1            action = agent.sample_action(state) # sample action next_state, reward, done, _ = env.step(action) # update env and return transitions agent.memory.push((state, action, reward,next_state, done)) # save transitions            state = next_state # update next state for env agent.update() # update agent ep_reward += reward  # if done: break steps.append(ep_step) rewards.append(ep_reward) if (i_ep + 1) % 10 == 0: print(f"回合：{i_ep+1}/{cfg['train_eps']}，奖励：{ep_reward:.2f}，Epislon: {agent.epsilon:.3f}") print("完成训练！") env.close() res_dic = {'episodes':range(len(rewards)),'rewards':rewards,'steps':steps} return res_dicdef test(cfg, env, agent): print("开始测试！") print(f"环境：{cfg['env_name']}，算法：{cfg['algo_name']}，设备：{cfg['device']}")    rewards = [] # record rewards for all episodes    steps = [] for i_ep in range(cfg['test_eps']): ep_reward = 0 # reward per episode ep_step = 0        state = env.reset() # reset and obtain initial state for _ in range(cfg['ep_max_steps']): ep_step+=1            action = agent.predict_action(state) # predict action next_state, reward, done, _ = env.step(action)             state = next_state  ep_reward += reward  if done: break steps.append(ep_step) rewards.append(ep_reward) print(f"回合：{i_ep+1}/{cfg['test_eps']}，奖励：{ep_reward:.2f}") print("完成测试！") env.close() return {'episodes':range(len(rewards)),'rewards':rewards,'steps':steps}

复制代码

3、定义环境

OpenAI Gym 中其实集成了很多强化学习环境，足够大家学习了，但是在做强化学习的应用中免不了要自己创建环境，比如在本项目中其实不太好找到 Qlearning 能学出来的环境，Qlearning 实在是太弱了，需要足够简单的环境才行，因此本项目写了一个环境，大家感兴趣的话可以看一下，一般环境接口最关键的部分即使 reset 和 step。

import gymimport paddleimport numpy as npimport randomimport osfrom parl.algorithms import DQNdef all_seed(env,seed = 1): ''' omnipotent seed for RL, attention the position of seed function, you'd better put it just following the env create function Args:        env (_type_):         seed (int, optional): _description_. Defaults to 1.    ''' print(f"seed = {seed}") env.seed(seed) # env config np.random.seed(seed) random.seed(seed) paddle.seed(seed)def env_agent_config(cfg): ''' create env and agent    '''    env = gym.make(cfg['env_name'])  if cfg['seed'] !=0: # set random seed all_seed(env,seed=cfg["seed"])  n_states = env.observation_space.shape[0] # print(hasattr(env.observation_space, 'n')) n_actions = env.action_space.n # action dimension print(f"n_states: {n_states}, n_actions: {n_actions}") cfg.update({"n_states":n_states,"n_actions":n_actions}) # update to cfg paramters    model = MLP(n_states,n_actions) algo = DQN(model, gamma=cfg['gamma'], lr=cfg['lr'])    memory = ReplayBuffer(cfg["memory_capacity"]) # replay buffer    agent = DQNAgent(algo,memory,cfg) # create agent return env, agent

复制代码

4、设置参数

到这里所有 qlearning 模块就算完成了，下面需要设置一些参数，方便大家“炼丹”，其中默认的是笔者已经调好的～。另外为了定义了一个画图函数，用来描述奖励的变化。

import argparseimport seaborn as snsimport matplotlib.pyplot as pltdef get_args(): """     """    parser = argparse.ArgumentParser(description="hyperparameters")  parser.add_argument('--algo_name',default='DQN',type=str,help="name of algorithm") parser.add_argument('--env_name',default='CartPole-v0',type=str,help="name of environment") parser.add_argument('--train_eps',default=200,type=int,help="episodes of training") # 训练的回合数 parser.add_argument('--test_eps',default=20,type=int,help="episodes of testing") # 测试的回合数 parser.add_argument('--ep_max_steps',default = 100000,type=int,help="steps per episode, much larger value can simulate infinite steps") parser.add_argument('--gamma',default=0.99,type=float,help="discounted factor") # 折扣因子 parser.add_argument('--epsilon_start',default=0.95,type=float,help="initial value of epsilon") #  e-greedy策略中初始epsilon parser.add_argument('--epsilon_end',default=0.01,type=float,help="final value of epsilon") # e-greedy策略中的终止epsilon parser.add_argument('--epsilon_decay',default=200,type=int,help="decay rate of epsilon") # e-greedy策略中epsilon的衰减率 parser.add_argument('--memory_capacity',default=200000,type=int) # replay memory的容量 parser.add_argument('--memory_warmup_size',default=200,type=int) # replay memory的预热容量 parser.add_argument('--batch_size',default=64,type=int,help="batch size of training") # 训练时每次使用的样本数 parser.add_argument('--targe_update_fre',default=200,type=int,help="frequency of target network update") # target network更新频率 parser.add_argument('--seed',default=10,type=int,help="seed")  parser.add_argument('--lr',default=0.0001,type=float,help="learning rate") parser.add_argument('--device',default='cpu',type=str,help="cpu or gpu")  args = parser.parse_args([])  args = {**vars(args)} # type(dict)          return argsdef smooth(data, weight=0.9):  '''用于平滑曲线，类似于Tensorboard中的smooth Args:        data (List):输入数据        weight (Float): 平滑权重，处于0-1之间，数值越高说明越平滑，一般取0.9    Returns:        smoothed (List): 平滑后的数据    '''    last = data[0] # First value in the plot (first timestep)    smoothed = list() for point in data: smoothed_val = last * weight + (1 - weight) * point  # 计算平滑值 smoothed.append(smoothed_val)         last = smoothed_val  return smootheddef plot_rewards(rewards,cfg,path=None,tag='train'): sns.set() plt.figure() # 创建一个图形实例，方便同时多画几个图 plt.title(f"{tag}ing curve on {cfg['device']} of {cfg['algo_name']} for {cfg['env_name']}") plt.xlabel('epsiodes') plt.plot(rewards, label='rewards') plt.plot(smooth(rewards), label='smoothed') plt.legend()

复制代码

5、训练

# 获取参数cfg = get_args() # 训练env, agent = env_agent_config(cfg)res_dic = train(cfg, env, agent)plot_rewards(res_dic['rewards'], cfg, tag="train") # 测试res_dic = test(cfg, env, agent)plot_rewards(res_dic['rewards'], cfg, tag="test") # 画出结果

复制代码

seed = 10n_states: 4, n_actions: 2开始训练！环境：CartPole-v0，算法：DQN，设备：cpu回合：10/200，奖励：10.00，Epislon: 0.062回合：20/200，奖励：85.00，Epislon: 0.014回合：30/200，奖励：41.00，Epislon: 0.011回合：40/200，奖励：31.00，Epislon: 0.010回合：50/200，奖励：22.00，Epislon: 0.010回合：60/200，奖励：10.00，Epislon: 0.010回合：70/200，奖励：10.00，Epislon: 0.010回合：80/200，奖励：22.00，Epislon: 0.010回合：90/200，奖励：30.00，Epislon: 0.010回合：100/200，奖励：20.00，Epislon: 0.010回合：110/200，奖励：15.00，Epislon: 0.010回合：120/200，奖励：45.00，Epislon: 0.010回合：130/200，奖励：73.00，Epislon: 0.010回合：140/200，奖励：180.00，Epislon: 0.010回合：150/200，奖励：167.00，Epislon: 0.010回合：160/200，奖励：200.00，Epislon: 0.010回合：170/200，奖励：165.00，Epislon: 0.010回合：180/200，奖励：200.00，Epislon: 0.010回合：190/200，奖励：200.00，Epislon: 0.010