maddpg结合优先经验回放和mpe环境 tensorflow代码

以下是使用TensorFlow实现MADDPG算法并结合优先经验回放和MPE环境的代码。代码中包含注释以帮助理解。

import numpy as np
import tensorflow as tf
import random

from mlagents.envs import UnityEnvironment
from mlagents.envs import BrainInfo

# 超参数
BUFFER_SIZE = int(1e6)  # 经验回放缓冲区大小
BATCH_SIZE = 128  # 批量大小
GAMMA = 0.99  # 折扣因子
TAU = 1e-3  # 目标网络更新速率
LR_ACTOR = 1e-3  # Actor网络学习率
LR_CRITIC = 1e-3  # Critic网络学习率
UPDATE_EVERY = 2  # 更新网络的时间步数
NUM_UPDATES = 10  # 每次更新网络的次数

# 神经网络模型
class Actor(tf.keras.Model):
    def __init__(self, state_size, action_size):
        super(Actor, self).__init__()
        self.fc1 = tf.keras.layers.Dense(256, activation='relu')
        self.fc2 = tf.keras.layers.Dense(128, activation='relu')
        self.fc3 = tf.keras.layers.Dense(action_size, activation='tanh')
        
    def call(self, state):
        x = self.fc1(state)
        x = self.fc2(x)
        x = self.fc3(x)
        return x
    
class Critic(tf.keras.Model):
    def __init__(self, state_size, action_size):
        super(Critic, self).__init__()
        self.fc1 = tf.keras.layers.Dense(256, activation='relu')
        self.fc2 = tf.keras.layers.Dense(128, activation='relu')
        self.fc3 = tf.keras.layers.Dense(1, activation=None)
        
        self.fc4 = tf.keras.layers.Dense(256, activation='relu')
        self.fc5 = tf.keras.layers.Dense(128, activation='relu')
        self.fc6 = tf.keras.layers.Dense(1, activation=None)
        
    def call(self, state, action):
        xs = tf.concat([state, action], axis=1)
        x1 = self.fc1(xs)
        x1 = self.fc2(x1)
        x1 = self.fc3(x1)
        
        x2 = self.fc4(xs)
        x2 = self.fc5(x2)
        x2 = self.fc6(x2)
        
        return x1, x2
    
# 优先经验回放类
class PrioritizedReplay:
    def __init__(self, buffer_size, batch_size):
        self.buffer_size = buffer_size
        self.batch_size = batch_size
        self.buffer = []
        self.priorities = np.zeros((buffer_size,), dtype=np.float32)
        self.pos = 0
        self.alpha = 0.5
        self.beta = 0.5
        self.beta_increment_per_sampling = 0.001
        
    def add(self, state, action, reward, next_state, done):
        max_priority = np.max(self.priorities) if self.buffer else 1.0
        experience = (state, action, reward, next_state, done)
        if len(self.buffer) < self.buffer_size:
            self.buffer.append(experience)
        else:
            self.buffer[self.pos] = experience
        self.priorities[self.pos] = max_priority
        self.pos = (self.pos + 1) % self.buffer_size
        
    def sample(self):
        if len(self.buffer) == self.buffer_size:
            priorities = self.priorities
        else:
            priorities = self.priorities[:self.pos]
        probs = priorities ** self.alpha
        probs /= probs.sum()
        indices = np.random.choice(len(self.buffer), self.batch_size, p=probs)
        samples = [self.buffer[idx] for idx in indices]
        total = len(self.buffer)
        weights = (total * probs[indices]) ** (-self.beta)
        weights /= weights.max()
        self.beta = np.min([1., self.beta + self.beta_increment_per_sampling])
        return indices, samples, weights
    
    def update_priorities(self, batch_indices, batch_priorities):
        for idx, priority in zip(batch_indices, batch_priorities):
            self.priorities[idx] = priority
            
# MADDPG算法类
class MADDPG:
    def __init__(self, state_size, action_size, num_agents):
        self.state_size = state_size
        self.action_size = action_size
        self.num_agents = num_agents
        
        self.actors = [Actor(state_size, action_size) for _ in range(num_agents)]
        self.critics = [Critic((state_size+action_size)*num_agents, 1) for _ in range(num_agents)]
        
        self.target_actors = [Actor(state_size, action_size) for _ in range(num_agents)]
        self.target_critics = [Critic((state_size+action_size)*num_agents, 1) for _ in range(num_agents)]
        
        for i in range(num_agents):
            self.target_actors[i].set_weights(self.actors[i].get_weights())
            self.target_critics[i].set_weights(self.critics[i].get_weights())
            
        self.buffer = PrioritizedReplay(BUFFER_SIZE, BATCH_SIZE)
        
        self.actor_optimizer = [tf.keras.optimizers.Adam(LR_ACTOR) for _ in range(num_agents)]
        self.critic_optimizer = [tf.keras.optimizers.Adam(LR_CRITIC) for _ in range(num_agents)]
        
        self.t_step = 0
        
    def act(self, obs):
        obs = np.array(obs)
        actions = []
        for i in range(self.num_agents):
            action = self.actors[i](obs[i][np.newaxis,:], training=False)
            actions.append(action.numpy())
        actions = np.concatenate(actions, axis=0)
        return actions
    
    def step(self, state, action, reward, next_state, done):
        self.buffer.add(state, action, reward, next_state, done)
        self.t_step = (self.t_step + 1) % UPDATE_EVERY
        if self.t_step == 0 and len(self.buffer.buffer) > BATCH_SIZE:
            for _ in range(NUM_UPDATES):
                indices, samples, weights = self.buffer.sample()
                self.learn(samples, weights)
                self.update_targets()
                self.buffer.update_priorities(indices, weights)
                
    def learn(self, samples, weights):
        states = np.array([sample[0] for sample in samples])
        actions = np.array([sample[1] for sample in samples])
        rewards = np.array([sample[2] for sample in samples])
        next_states = np.array([sample[3] for sample in samples])
        dones = np.array([sample[4] for sample in samples])
        
        for i in range(self.num_agents):
            # 计算Q值
            with tf.GradientTape(persistent=True) as tape:
                target_actions = [self.target_actors[j](next_states[j][np.newaxis,:], training=False) for j in range(self.num_agents)]
                target_actions = np.concatenate(target_actions, axis=0)
                target_qs = self.target_critics[i]((next_states.reshape(-1, self.state_size*self.num_agents), target_actions))
                target_qs = target_qs.numpy().reshape(-1, self.num_agents)
                q_targets = rewards[:,i][:,np.newaxis] + (GAMMA * target_qs * (1 - dones[:,i][:,np.newaxis]))
                critic_qs = self.critics[i]((states.reshape(-1, self.state_size*self.num_agents), actions.reshape(-1, self.action_size*self.num_agents)))
                critic_loss = tf.reduce_mean(weights * (q_targets - critic_qs)**2)
            critic_grads = tape.gradient(critic_loss, self.critics[i].trainable_variables)
            self.critic_optimizer[i].apply_gradients(zip(critic_grads, self.critics[i].trainable_variables))
            
            # 计算Actor梯度
            with tf.GradientTape() as tape:
                actor_actions = [self.actors[j](states[:,j,:], training=False) if j == i else self.actors[j](states[:,j,:], training=True) for j in range(self.num_agents)]
                actor_actions = np.concatenate(actor_actions, axis=0)
                actor_loss = -tf.reduce_mean(self.critics[i]((states.reshape(-1, self.state_size*self.num_agents), actor_actions)))
            actor_grads = tape.gradient(actor_loss, self.actors[i].trainable_variables)
            self.actor_optimizer[i].apply_gradients(zip(actor_grads, self.actors[i].trainable_variables))
            
    def update_targets(self):
        for i in range(self.num_agents):
            self.target_actors[i].set_weights(TAU*np.array(self.actors[i].get_weights())+(1-TAU)*np.array(self.target_actors[i].get_weights()))
            self.target_critics[i].set_weights(TAU*np.array(self.critics[i].get_weights())+(1-TAU)*np.array(self.target_critics[i].get_weights()))

# 环境
env_name = "MPE/3DBall"
env = UnityEnvironment(file_name=env_name)
brain_name = env.brain_names[0]
brain = env.brains[brain_name]
env_info = env.reset()[brain_name]

state_size = env_info.vector_observations.shape[1]
action_size = brain.vector_action_space_size
num_agents = len(env_info.agents)

maddpg = MADDPG(state_size, action_size, num_agents)

scores = []
scores_window = deque(maxlen=100)
for i_episode in range(10000):
    env_info = env.reset()[brain_name]
    obs = env_info.vector_observations
    score = np.zeros(num_agents)
    while True:
        actions = maddpg.act(obs)
        env_info = env.step(actions)[brain_name]
        next_obs = env_info.vector_observations
        rewards = env_info.rewards
        dones = env_info.local_done
        maddpg.step(obs, actions, rewards, next_obs, dones)
        obs = next_obs
        score += rewards
        if np.any(dones):
            break
    scores_window.append(np.max(score))
    scores.append(np.max(score))
    print('\rEpisode {}\tAverage Score: {:.2f}'.format(i_episode, np.mean(scores_window)), end="")
    if i_episode % 100 == 0:
        print('\rEpisode {}\tAverage Score: {:.2f}'.format(i_episode, np.mean(scores_window)))
    if np.mean(scores_window)>=0.5:
        print('\nEnvironment solved in {:d} episodes!\tAverage Score: {:.2f}'.format(i_episode, np.mean(scores_window)))
        break

env.close()