使用.detach的Pytorch DQN，DDQN造成了非常大的损失呈指数增长并且根本不学习

如何解决使用.detach的Pytorch DQN，DDQN造成了非常大的损失呈指数增长并且根本不学习

这是我对CartPole-v0的DQN和DDQN的实现，我认为是正确的。

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,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,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,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,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,hidden_dim=16).double()
      self.target_network = NETWORK(input_dim=4,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,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,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,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,info = env.step(action)
        episodic_reward += reward      
        sign = 1 if done else 0
        agent.train(state,action,next_state,sign)
        state = next_state        
    print(f'episode: {i},reward: {episodic_reward}')  
    if i % 20 == 0:
        eval_reward = eval_policy(agent,eval_episodes=50)
        if eval_reward >= 195:
            print("Problem solved in {} episodes".format(i + 1))
            break

问题是我的DQN网络没有进行训练，并且在损耗计算中使用target.detach（）导致损耗成指数增长。如果我不使用.detach（），DQN对象将进行训练，但是我认为这不是正确的方法。对于DDQN，我的网络始终不训练。任何人都可以针对哪里可能出问题提供一些建议吗？

解决方法

所以实现中的一个错误是，您永远不要将情节的结尾添加到重播缓冲区中。在火车功能中，如果sign == 1（情节结束），则返回。删除返回值，并通过（1-dones）* ...调整目标计算，以防您采样情节结尾的过渡。情节结尾之所以重要的原因是，这是唯一的经历，即无法通过引导来逼近目标。然后DQN训练。为了重现性，我使用0.99的贴现率和2020年的种子折扣（用于火炬，麻木和体育馆环境）。经过241次训练，我获得了199.100的奖励。

希望有帮助，顺便说一句，代码很可读。