强化学习——Actor Critic Method

本文介绍CartPole-V0环境中Actor-Critic方法的实现。该方法含Actor和Critic两个网络，前者输出动作概率，后者估计未来回报。训练时，通过交互收集数据，计算回报和优势，分别更新两个网络。实验显示，训练后代理能长时间保持杆子平衡，体现了该方法结合策略与值函数逼近、单步更新、高效利用数据的优势。

强化学习——Actor Critic Method

一、介绍

这个脚本展示了CartPole-V0环境中Actor-Critic方法的一个实现。

Actor Critic Method（演员--评论家算法）

当代理在环境中执行操作和移动时，它将学习将观察到的环境状态映射到两个可能的输出：

CartPole-V0

在无摩擦的轨道上，一根杆子系在一辆推车上。agent（代理）必须施加力才能移动推车。每走一步，杆子就保持直立，这是奖励。因此，agent（代理）必须学会防止杆子掉下来。

二、环境配置

本教程基于Paddle 2.0 编写，如果您的环境不是本版本，请先参考正式安装 Paddle 2.0 。

import gym, osfrom itertools import countimport paddleimport paddle.nn as nnimport paddle.optimizer as optimimport paddle.nn.functional as Ffrom paddle.distribution import Categorical

三、实施演员-评论家网络

这个网络学习两个功能：

device = paddle.get_device()env = gym.make("CartPole-v0")  ### 或者 env = gym.make("CartPole-v0").unwrapped 开启无锁定环境训练state_size = env.observation_space.shape[0]action_size = env.action_space.nlr = 0.001##定义“演员”网络class Actor(nn.Layer):    def __init__(self, state_size, action_size):        super(Actor, self).__init__()        self.state_size = state_size        self.action_size = action_size        self.linear1 = nn.Linear(self.state_size, 128)        self.linear2 = nn.Linear(128, 256)        self.linear3 = nn.Linear(256, self.action_size)    def forward(self, state):        output = F.relu(self.linear1(state))        output = F.relu(self.linear2(output))        output = self.linear3(output)        distribution = Categorical(F.softmax(output, axis=-1))        return distribution##定义“评论家”网络class Critic(nn.Layer):    def __init__(self, state_size, action_size):        super(Critic, self).__init__()        self.state_size = state_size        self.action_size = action_size        self.linear1 = nn.Linear(self.state_size, 128)        self.linear2 = nn.Linear(128, 256)        self.linear3 = nn.Linear(256, 1)    def forward(self, state):        output = F.relu(self.linear1(state))        output = F.relu(self.linear2(output))        value = self.linear3(output)        return value

四、训练模型

def compute_returns(next_value, rewards, masks, gamma=0.99):    R = next_value    returns = []    for step in reversed(range(len(rewards))):        R = rewards[step] + gamma * R * masks[step]        returns.insert(0, R)    return returns## 定义训练过程def trainIters(actor, critic, n_iters):    optimizerA = optim.Adam(lr, parameters=actor.parameters())    optimizerC = optim.Adam(lr, parameters=critic.parameters())    for iter in range(n_iters):        state = env.reset()        log_probs = []        values = []        rewards = []        masks = []        entropy = 0        env.reset()        for i in count():            # env.render()            state = paddle.to_tensor(state,dtype="float32",place=device)            dist, value = actor(state), critic(state)            action = dist.sample([1])            next_state, reward, done, _ = env.step(action.cpu().squeeze(0).numpy()) #env.step(action.cpu().squeeze(0).numpy())            log_prob = dist.log_prob(action);            entropy += dist.entropy().mean()            log_probs.append(log_prob)            values.append(value)            rewards.append(paddle.to_tensor([reward], dtype="float32", place=device))            masks.append(paddle.to_tensor([1-done], dtype="float32", place=device))            state = next_state            if done:                if iter % 10 == 0:                    print('Iteration: {}, Score: {}'.format(iter, i))                break        next_state = paddle.to_tensor(next_state, dtype="float32", place=device)        next_value = critic(next_state)        returns = compute_returns(next_value, rewards, masks)        log_probs = paddle.concat(log_probs)        returns = paddle.concat(returns).detach()        values = paddle.concat(values)        advantage = returns - values        actor_loss = -(log_probs * advantage.detach()).mean()        critic_loss = advantage.pow(2).mean()        optimizerA.clear_grad()        optimizerC.clear_grad()        actor_loss.backward()        critic_loss.backward()        optimizerA.step()        optimizerC.step()    paddle.save(actor.state_dict(), 'model/actor.pdparams')    paddle.save(critic.state_dict(), 'model/critic.pdparams')    env.close()if __name__ == '__main__':    if os.path.exists('model/actor.pdparams'):        actor = Actor(state_size, action_size)        model_state_dict  = paddle.load('model/actor.pdparams')        actor.set_state_dict(model_state_dict )        print('Actor Model loaded')    else:        actor = Actor(state_size, action_size)    if os.path.exists('model/critic.pdparams'):        critic = Critic(state_size, action_size)        model_state_dict  = paddle.load('model/critic.pdparams')        critic.set_state_dict(model_state_dict )        print('Critic Model loaded')    else:        critic = Critic(state_size, action_size)    trainIters(actor, critic, n_iters=201)

Iteration: 0, Score: 19Iteration: 10, Score: 59Iteration: 20, Score: 24Iteration: 30, Score: 33Iteration: 40, Score: 39Iteration: 50, Score: 62Iteration: 60, Score: 44Iteration: 70, Score: 59Iteration: 80, Score: 21Iteration: 90, Score: 85Iteration: 100, Score: 152Iteration: 110, Score: 103Iteration: 120, Score: 69Iteration: 130, Score: 170Iteration: 140, Score: 199Iteration: 150, Score: 197Iteration: 160, Score: 199Iteration: 170, Score: 199Iteration: 180, Score: 163Iteration: 190, Score: 199Iteration: 200, Score: 199

开启无锁定环境训练

200分就是上限，还不能体现模型的能力，现在咱们开启无锁定环境的训练,将env = gym.make("CartPole-v0").unwrapped ,可以适当提高训练轮次。 （所有代码均已封装到train.py文件中，并用vdl记录）

!python train.py

W0303 17:18:01.972481  3940 device_context.cc:362] Please NOTE: device: 0, GPU Compute Capability: 7.0, Driver API Version: 10.1, Runtime API Version: 10.1W0303 17:18:01.977756  3940 device_context.cc:372] device: 0, cuDNN Version: 7.6.Actor Model loadedCritic Model loadedIteration: 0, Score: 325

这是用vdl记录的Agent可以连续保持平衡的轮数的记录,最大值已经高达131540了，这是一个惊人的数字。

五、效果展示

在训练的早期         

在训练的后期         

六、总结

你是一名 AI 行业编辑，请围绕下面这条热点输出一份资讯解读：
热点：强化学习——Actor Critic Method要求：
1. 先用一句话解释这条热点在讲什么
2. 再总结它为什么重要
3. 说明会影响哪些 AI 产品或内容方向
4. 最后给出 3 个适合资讯站使用的标题