使用 .detach() 的 Pytorch DQN、DDQN 导致非常大的损失(呈指数增长)并且根本不学习

Question

这是我为 CartPole-v0 实现的 DQN 和 DDQN，我认为这是正确的。

import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
import gym
import torch.optim as optim
import random
import os
import time


class NETWORK(torch.nn.Module):
    def __init__(self, input_dim: int, output_dim: int, hidden_dim: int) -> None:

        super(NETWORK, self).__init__()

        self.layer1 = torch.nn.Sequential(
            torch.nn.Linear(input_dim, hidden_dim),
            torch.nn.ReLU()
        )

        self.layer2 = torch.nn.Sequential(
            torch.nn.Linear(hidden_dim, hidden_dim),
            torch.nn.ReLU()
        )

        self.final = torch.nn.Linear(hidden_dim, output_dim)

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        x = self.layer1(x)
        x = self.layer2(x)
        x = self.final(x)

        return x

class ReplayBuffer(object):
    def __init__(self, capacity=50000):
        self.capacity = capacity
        self.memory = []
        self.position = 0

    def push(self, s0, a0, r, s1):
        if len(self.memory) < self.capacity:
            self.memory.append(None)
        self.memory[self.position] = (s0, a0, r, s1)
        self.position = (self.position + 1) % self.capacity

    def sample(self, batch_size=64):
        return random.sample(self.memory, batch_size)

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

class DQN(object):
    def __init__(self):
      self.state_dim = 4
      self.action_dim = 2
      self.lr = 0.001
      self.discount_factor = 0.99
      self.epsilon = 1
      self.epsilon_decay = 0.95
      self.num_train = 0
      self.num_train_episodes = 0
      self.batch_size = 64

      self.predict_network = NETWORK(input_dim=4, output_dim=2, hidden_dim=16).double()

      self.memory = ReplayBuffer(capacity=50000)
      self.optimizer = torch.optim.Adam(self.predict_network.parameters(), lr=self.lr)
      self.loss = 0

    def select_action(self, states: np.ndarray) -> int:
      if np.random.uniform(0, 1) < self.epsilon:
        return np.random.choice(self.action_dim)
      else:
        states = torch.from_numpy(states).unsqueeze_(dim=0)
        with torch.no_grad():
          Q_values = self.predict_network(states)
          action = torch.argmax(Q_values).item()
        return action

    def policy(self, states: np.ndarray) -> int:
      states = torch.from_numpy(states).unsqueeze_(dim=0)
      with torch.no_grad():
        Q_values = self.predict_network(states)
        action = torch.argmax(Q_values).item()
      return action

    def train(self, s0, a0, r, s1, sign):
      if sign == 1:
        self.num_train_episodes += 1
        if self.epsilon > 0.01:
          self.epsilon = max(self.epsilon * self.epsilon_decay, 0.01)
        return

      self.num_train += 1
      self.memory.push(s0, a0, r, s1)
      if len(self.memory) < self.batch_size:
        return
      
      batch = self.memory.sample(self.batch_size)
      state_batch = torch.from_numpy(np.stack([b[0] for b in batch]))
      action_batch = torch.from_numpy(np.stack([b[1] for b in batch]))
      reward_batch = torch.from_numpy(np.stack([b[2] for b in batch]))
      next_state_batch = torch.from_numpy(np.stack([b[3] for b in batch]))

      Q_values = self.predict_network(state_batch)[torch.arange(self.batch_size), action_batch]
      
      next_state_Q_values = self.predict_network(next_state_batch).max(dim=1)[0]
      
      Q_targets = self.discount_factor * next_state_Q_values + reward_batch
      
      loss = F.mse_loss(Q_values, Q_targets.detach())
      
      self.optimizer.zero_grad()
      loss.backward()
      self.optimizer.step()

      self.loss = loss.data.item()

class DDQN(object):
    def __init__(self):
      self.state_dim = 4
      self.action_dim = 2
      self.lr = 0.001
      self.discount_factor = 0.9
      self.epsilon = 1
      self.epsilon_decay = 0.95
      self.num_train = 0
      self.num_train_episodes = 0
      self.batch_size = 64

      self.predict_network = NETWORK(input_dim=4, output_dim=2, hidden_dim=16).double()
      self.target_network = NETWORK(input_dim=4, output_dim=2, hidden_dim=16).double()
      self.target_network.load_state_dict(self.predict_network.state_dict())
      self.target_network.eval()

      self.memory = ReplayBuffer(capacity=50000)
      self.optimizer = torch.optim.Adam(self.predict_network.parameters(), lr=self.lr)

      self.loss = 0

    def select_action(self, states: np.ndarray) -> int:
      if np.random.uniform(0, 1) < self.epsilon:
        return np.random.choice(self.action_dim)
      else:
        states = torch.from_numpy(states).unsqueeze_(dim=0)
        with torch.no_grad():
          Q_values = self.predict_network(states)
        action = torch.argmax(Q_values).item()
        return action

    def policy(self, states: np.ndarray) -> int:
      states = torch.from_numpy(states).unsqueeze_(dim=0)
      with torch.no_grad():
        Q_values = self.predict_network(states)
        action = torch.argmax(Q_values).item()
      return action

    def train(self, s0, a0, r, s1, sign):
      if sign == 1:
        self.num_train_episodes += 1
        if self.num_train_episodes % 2 == 0:
          self.target_network.load_state_dict(self.predict_network.state_dict())
          self.target_network.eval()
          
          if self.epsilon > 0.01:
            self.epsilon = max(self.epsilon * self.epsilon_decay, 0.01)
        return

      self.num_train += 1
      self.memory.push(s0, a0, r, s1)
      if len(self.memory) < self.batch_size:
        return
      batch = self.memory.sample(self.batch_size)
      state_batch = torch.from_numpy(np.stack([b[0] for b in batch]))
      action_batch = torch.from_numpy(np.stack([b[1] for b in batch]))
      reward_batch = torch.from_numpy(np.stack([b[2] for b in batch]))
      next_state_batch = torch.from_numpy(np.stack([b[3] for b in batch]))
      
      Q_values = self.predict_network(state_batch)[torch.arange(self.batch_size), action_batch]
      
      next_state_action_batch = torch.argmax(self.predict_network(next_state_batch), dim=1)
      
      next_state_Q_values = self.target_network(next_state_batch)[torch.arange(self.batch_size), next_state_action_batch]
      
      Q_targets = self.discount_factor * next_state_Q_values + reward_batch
      
      loss = F.smooth_l1_loss(Q_values, Q_targets.detach())
      self.optimizer.zero_grad()
      loss.backward()
      self.optimizer.step()

      self.loss = loss.data.item()

我使用以下方法来评估和训练我的 DQN 和 DDQN。

def eval_policy(agent, env_name, eval_episodes=10):
    eval_env = gym.make(env_name)
    avg_reward = 0.
    for _ in range(eval_episodes):
        state, done = eval_env.reset(), False
        while not done:
            action = agent.policy(state)
            state, reward, done, _ = eval_env.step(action)
            avg_reward += reward
    avg_reward /= eval_episodes
    print("---------------------------------------")
    print(f"Evaluation over {eval_episodes} episodes: {avg_reward:.3f}")
    print("---------------------------------------")
    return avg_reward


env_name = 'CartPole-v0'
env = gym.make(env_name)
    
agent = DQN() # agent = DDQN()

for i in range(1000):
    state, done = env.reset(), False
    episodic_reward = 0
    while not done:
        action = agent.select_action(np.squeeze(state))
        next_state, reward, done, info = env.step(action)
        episodic_reward += reward      
        sign = 1 if done else 0
        agent.train(state, action, reward, next_state, sign)
        state = next_state        
    print(f'episode: {i}, reward: {episodic_reward}')  
    if i % 20 == 0:
        eval_reward = eval_policy(agent, env_name, eval_episodes=50)
        if eval_reward >= 195:
            print("Problem solved in {} episodes".format(i + 1))
            break

问题是我的 DQN 网络没有训练，并且在损失计算中使用 target.detach() 使损失呈指数增长。如果我不使用 .detach()，DQN 对象将进行训练，但我认为这不是正确的方法。对于 DDQN，我的网络总是不训练。任何人都可以就可能出错的地方提供一些建议吗？

Answer 1

所以您实现中的一个错误是您从未将一集的结尾添加到您的重播缓冲区。在您的火车功能中，如果 sign==1 (剧集结束)，您将返回。删除该返回并通过 (1-dones)*... 调整目标计算，以防您对剧集结尾的过渡进行采样。剧集结尾之所以重要，是因为它是唯一的体验，目标不是通过自举逼近。然后DQN训练。为了再现性，我使用了 0.99 的折扣率和 2020 种子(用于 torch、numpy 和健身房环境)。经过 241 次训练后，我获得了 199.100 的奖励。

希望对您有所帮助，顺便说一下，代码非常易读。