强化学习 - Twin Delayed DDPG (TD3)

什么是机器学习

Twin Delayed DDPG (TD3) 是一种用于解决连续动作空间的强化学习问题的算法，是 Deep Deterministic Policy Gradient (DDPG) 的改进版本。TD3引入了一些技巧，例如双Q网络（Twin Q-networks）和延迟更新，以提高算法的性能和稳定性。

以下是一个使用 Python 和 TensorFlow/Keras 实现简单的 TD3 的示例。在这个例子中，我们将使用 OpenAI Gym 的 Pendulum 环境。

import numpy as np
import tensorflow as tf
from tensorflow.keras.models import Model
from tensorflow.keras.layers import Dense, Input
from tensorflow.keras.optimizers import Adam
import gym

# 定义TD3 Agent
class TD3Agent:
    def __init__(self, state_size, action_size):
        self.state_size = state_size
        self.action_size = action_size
        self.gamma = 0.99  # 折扣因子
        self.tau = 0.005  # 软更新参数
        self.actor_lr = 0.001
        self.critic_lr = 0.001
        self.policy_noise = 0.2
        self.policy_noise_clip = 0.5
        self.exploration_noise = 0.1
        self.buffer_size = 1000000
        self.batch_size = 100
        self.buffer = []

        # 构建演员（Actor）网络和目标演员网络
        self.actor = self.build_actor()
        self.target_actor = self.build_actor()
        self.target_actor.set_weights(self.actor.get_weights())

        # 构建两个评论家（Critic）网络和目标评论家网络
        self.critic_1 = self.build_critic()
        self.target_critic_1 = self.build_critic()
        self.target_critic_1.set_weights(self.critic_1.get_weights())

        self.critic_2 = self.build_critic()
        self.target_critic_2 = self.build_critic()
        self.target_critic_2.set_weights(self.critic_2.get_weights())

    def build_actor(self):
        state_input = Input(shape=(self.state_size,))
        dense1 = Dense(400, activation='relu')(state_input)
        dense2 = Dense(300, activation='relu')(dense1)
        output = Dense(self.action_size, activation='tanh')(dense2)
        model = Model(inputs=state_input, outputs=output)
        model.compile(loss='mse', optimizer=Adam(lr=self.actor_lr))
        return model

    def build_critic(self):
        state_input = Input(shape=(self.state_size,))
        action_input = Input(shape=(self.action_size,))
        concat = tf.keras.layers.concatenate([state_input, action_input])
        dense1 = Dense(400, activation='relu')(concat)
        dense2 = Dense(300, activation='relu')(dense1)
        output = Dense(1, activation='linear')(dense2)
        model = Model(inputs=[state_input, action_input], outputs=output)
        model.compile(loss='mse', optimizer=Adam(lr=self.critic_lr))
        return model

    def get_action(self, state):
        state = np.reshape(state, [1, self.state_size])
        action = self.actor.predict(state)[0]
        action = np.clip(action + np.random.normal(0, self.exploration_noise, self.action_size), -1, 1)
        return action

    def train(self):
        if len(self.buffer) < self.batch_size:
            return

        batch = np.random.choice(self.buffer, self.batch_size, replace=False)
        states, actions, rewards, next_states, dones = zip(*batch)

        states = np.vstack(states)
        actions = np.vstack(actions)
        rewards = np.vstack(rewards)
        next_states = np.vstack(next_states)
        dones = np.vstack(dones)

        next_actions = self.target_actor.predict(next_states) + np.clip(np.random.normal(0, self.policy_noise, self.action_size), -self.policy_noise_clip, self.policy_noise_clip)
        next_actions = np.clip(next_actions, -1, 1)

        target_q_values = np.minimum(
            self.target_critic_1.predict([next_states, next_actions]),
            self.target_critic_2.predict([next_states, next_actions])
        )
        target_values = rewards + self.gamma * (1 - dones) * target_q_values

        self.critic_1.train_on_batch([states, actions], target_values)
        self.critic_2.train_on_batch([states, actions], target_values)

        actor_gradients = np.reshape(self.critic_1.gradient(states + self.actor.predict(states)), [-1, self.action_size])
        actor_gradients = actor_gradients / self.batch_size

        self.actor.train_on_batch(states, actor_gradients)

        self.soft_update_target_networks()

    def soft_update_target_networks(self):
        actor_weights = np.array(self.actor.get_weights())
        target_actor_weights = np.array(self.target_actor.get_weights())
        self.target_actor.set_weights(self.tau * actor_weights + (1 - self.tau) * target_actor_weights)

        critic_1_weights = np.array(self.critic_1.get_weights())
        target_critic_1_weights = np.array(self.target_critic_1.get_weights())
        self.target_critic_1.set_weights(self.tau * critic_1_weights + (1 - self.tau) * target_critic_1_weights)

        critic_2_weights = np.array(self.critic_2.get_weights())
        target_critic_2_weights = np.array(self.target_critic_2.get_weights())
        self.target_critic_2.set_weights(self.tau * critic_2_weights + (1 - self.tau) * target_critic_2_weights)

    def store_experience(self, state, action, reward, next_state, done):
        self.buffer.append((state, action, reward, next_state, done))
        if len(self.buffer) > self.buffer_size:
            self.buffer.pop(0)

# 初始化环境和Agent
env = gym.make('Pendulum-v0')
state_size = env.observation_space.shape[0]
action_size = env.action_space.shape[0]
agent = TD3Agent(state_size, action_size)

# 训练TD3 Agent
num_episodes = 500
for episode in range(num_episodes):
    state = env.reset()
    total_reward = 0
    for time in range(500):  # 限制每个episode的步数，防止无限循环
        # env.render()  # 如果想可视化训练过程，可以取消注释此行
        action = agent.get_action(state)
        next_state, reward, done, _ = env.step(action)
        total_reward += reward
        agent.store_experience(state, action, reward, next_state, done)
        agent.train()
        state = next_state
        if done:
            print("Episode: {}, Total Reward: {}".format(episode + 1, total_reward))
            break

# 关闭环境
env.close()

在这个例子中，我们定义了一个简单的TD3 Agent，包括演员（Actor）和两个评论家（Critic）神经网络。在训练过程中，我们使用了两个评论家网络和一些技巧来提高稳定性，并进行了软更新。请注意，TD3算法的实现可能因问题的复杂性而有所不同，可能需要更多的技术和调整，如归一化奖励、使用更复杂的神经网络结构等。