一、倒立摆问题介绍

Agent 必须在两个动作之间做出决定 - 向左或向右移动推车 - 以使连接到它的杆保持直立。
在这里插入图片描述

二、深度Q网络简介

在这里插入图片描述

上图所示为一般的深度 $\mathrm{Q}$ 网络算法。

深度 $\mathrm{Q}$ 网络算法是这样的，我们初始化两个网络：估计网络 $Q$ 和目标网络 $\hat{Q} ， \hat{Q}$ 就等于 $Q$ ，一开始目标网络 $\hat{Q}$ 与原来的 $Q$ 网络是一样的。

在每一个回合中，我们用演员与环境交互，在每一次交互的过程中，都会得到一个状态 $s_t$ ，会采取某一个动作 $a_{t 。}$ 怎么知道采取哪一个动作 $a_t$ 呢? 我们就根据现在的 Q函数，但是要有探索的机制。比如我们用玻尔兹曼探索或是 $\varepsilon$ -贪心探索，接下来得到奖励 $r_t$ ，进入状态 $s_{t+1}$ 。

所以现在收集到一笔数据 $\left(s_t 、 a_t 、 r_t 、 s_{t+1}\right)$ ，我们将其放到回放缓冲区里面。如果回放缓冲区满了，我们就把一些旧的数据丢掉。

接下来我们就从回放缓冲区里面去采样数据，采样到的是 $\left(s_i 、 a_i 、 r_i 、 s_{i+1}\right)$ 。这笔数据与刚放进去的不一定是同一笔，我们可能抽到旧的。要注意的是， 我们采样出来不是一笔数据，采样出来的是一个批量的数据，采样一些经验出来。

接下来就是计算目标。假设我们采样出一笔数据，根据这笔数据去计算目标。目标要用目标网络 $\hat{Q}$ 来计算。目标是:
$y=r_i+\max _a \hat{Q}\left(s_{i+1}, a\right)$
其中， $a$ 是让 $\hat{Q}$ 值最大的动作。因为我们在状态 $s_{i+1}$ 会采取的动作 $a$ 就是可以让 $\hat{Q}$ 值最大的那一个动作。接下来我们要更新 $\mathrm{Q}$ 值，就把它当作一个回归问题。我们希望 $Q\left(s_i, a_i\right)$ 与目标越接近越好。

假设已经更新了一定的次数，比如 $C$ 次，设 $C = 100$ ，那我们就把 $\hat{Q}$ 设成 $Q$ ，这就是深度Q网络算法。

三、详细资料

关于更加详细的深度Q网络的介绍，请看我之前发的博客：【EasyRL学习笔记】第六章 DQN 深度Q网络（基本概念）

在学习深度Q网络前你最好能了解以下知识点：

全连接神经网络
神经网络求解分类问题
神经网络基本工作原理
Q-Learning算法

四、Python代码实战

4.1 运行前配置

准备好一个RL_Utils.py文件，文件内容可以从我的一篇里博客获取：【RL工具类】强化学习常用函数工具类（Python代码）

这一步很重要，后面需要引入该RL_Utils.py文件

在这里插入图片描述

4.2 主要代码

import argparse
import datetime
import time
import math
import torch.optim as optim
import gym
from torch import nn# 这里需要改成自己的RL_Utils.py文件的路径
from Python.ReinforcementLearning.EasyRL.RL_Utils import *# Q网络（3层全连接网络）
class MLP(nn.Module):def __init__(self, input_dim, output_dim, hidden_dim=128):""" 初始化q网络，为全连接网络input_dim: 输入的特征数即环境的状态维度output_dim: 输出的动作维度"""super(MLP, self).__init__()self.fc1 = nn.Linear(input_dim, hidden_dim)  # 输入层self.fc2 = nn.Linear(hidden_dim, hidden_dim)  # 隐藏层self.fc3 = nn.Linear(hidden_dim, output_dim)  # 输出层def forward(self, x):# 各层对应的激活函数x = torch.relu(self.fc1(x))x = torch.relu(self.fc2(x))return self.fc3(x)# 经验回放缓存区
class ReplayBuffer:def __init__(self, capacity):self.capacity = capacity  # 经验回放的容量self.buffer = []  # 缓冲区self.position = 0def push(self, state, action, reward, next_state, done):''' 缓冲区是一个队列，容量超出时去掉开始存入的转移(transition)'''if len(self.buffer) < self.capacity:self.buffer.append(None)self.buffer[self.position] = (state, action, reward, next_state, done)self.position = (self.position + 1) % self.capacitydef sample(self, batch_size):batch = random.sample(self.buffer, batch_size)  # 随机采出小批量转移state, action, reward, next_state, done = zip(*batch)  # 解压成状态，动作等return state, action, reward, next_state, donedef __len__(self):''' 返回当前存储的量'''return len(self.buffer)# DQN智能体对象
class DQN:def __init__(self, model, memory, cfg):self.n_actions = cfg['n_actions']self.device = torch.device(cfg['device'])self.gamma = cfg['gamma']## e-greedy 探索策略参数self.sample_count = 0  # 采样次数self.epsilon = cfg['epsilon_start']self.sample_count = 0self.epsilon_start = cfg['epsilon_start']self.epsilon_end = cfg['epsilon_end']self.epsilon_decay = cfg['epsilon_decay']self.batch_size = cfg['batch_size']self.policy_net = model.to(self.device)self.target_net = model.to(self.device)# 初始化的时候，目标Q网络和估计Q网络相等，将策略网络的参数复制给目标网络self.target_net.load_state_dict(self.policy_net.state_dict())self.optimizer = optim.Adam(self.policy_net.parameters(), lr=cfg['lr'])self.memory = memoryself.update_flag = False# 训练过程采样：e-greedy policydef sample_action(self, state):self.sample_count += 1self.epsilon = self.epsilon_end + (self.epsilon_start - self.epsilon_end) * \math.exp(-1. * self.sample_count / self.epsilon_decay)if random.random() > self.epsilon:return self.predict_action(state)else:action = random.randrange(self.n_actions)return action# 测试过程：以最大Q值选取动作def predict_action(self, state):with torch.no_grad():state = torch.tensor(state, device=self.device, dtype=torch.float32).unsqueeze(dim=0)q_values = self.policy_net(state)action = q_values.max(1)[1].item()return actiondef update(self):# 当经验缓存区没有满的时候，不进行更新if len(self.memory) < self.batch_size:returnelse:if not self.update_flag:print("Begin to update!")self.update_flag = True# 从经验缓存区随机取出一个batch的数据state_batch, action_batch, reward_batch, next_state_batch, done_batch = self.memory.sample(self.batch_size)# 将数据转化成Tensor格式state_batch = torch.tensor(np.array(state_batch), device=self.device,dtype=torch.float)  # shape(batchsize,n_states)action_batch = torch.tensor(action_batch, device=self.device).unsqueeze(1)  # shape(batchsize,1)reward_batch = torch.tensor(reward_batch, device=self.device, dtype=torch.float).unsqueeze(1)  # shape(batchsize,1)next_state_batch = torch.tensor(np.array(next_state_batch), device=self.device,dtype=torch.float)  # shape(batchsize,n_states)done_batch = torch.tensor(np.float32(done_batch), device=self.device).unsqueeze(1)  # shape(batchsize,1)# 计算Q估计q_value_batch = self.policy_net(state_batch).gather(dim=1,index=action_batch)  # shape(batchsize,1),requires_grad=Truenext_max_q_value_batch = self.target_net(next_state_batch).max(1)[0].detach().unsqueeze(1)# 计算Q现实expected_q_value_batch = reward_batch + self.gamma * next_max_q_value_batch * (1 - done_batch)# 计算损失函数MSE（Q估计，Q现实）loss = nn.MSELoss()(q_value_batch, expected_q_value_batch)# 梯度下降self.optimizer.zero_grad()loss.backward()# 限制梯度的范围，以避免梯度爆炸for param in self.policy_net.parameters():param.grad.data.clamp_(-1, 1)self.optimizer.step()def save_model(self, path):Path(path).mkdir(parents=True, exist_ok=True)torch.save(self.target_net.state_dict(), f"{path}/checkpoint.pt")def load_model(self, path):self.target_net.load_state_dict(torch.load(f"{path}/checkpoint.pt"))for target_param, param in zip(self.target_net.parameters(), self.policy_net.parameters()):param.data.copy_(target_param.data)# 训练函数
def train(arg_dict, env, agent):# 开始计时startTime = time.time()print(f"环境名: {arg_dict['env_name']}, 算法名: {arg_dict['algo_name']}, Device: {arg_dict['device']}")print("开始训练智能体......")rewards = []steps = []for i_ep in range(arg_dict["train_eps"]):ep_reward = 0ep_step = 0state = env.reset()for _ in range(arg_dict['ep_max_steps']):# 画图if arg_dict['train_render']:env.render()ep_step += 1action = agent.sample_action(state)next_state, reward, done, _ = env.step(action)agent.memory.push(state, action, reward,next_state, done)state = next_stateagent.update()ep_reward += rewardif done:break# 目标网络更新if (i_ep + 1) % arg_dict["target_update"] == 0:agent.target_net.load_state_dict(agent.policy_net.state_dict())steps.append(ep_step)rewards.append(ep_reward)if (i_ep + 1) % 10 == 0:print(f'Episode: {i_ep + 1}/{arg_dict["train_eps"]}, Reward: {ep_reward:.2f}: Epislon: {agent.epsilon:.3f}')print('训练结束 , 用时: ' + str(time.time() - startTime) + " s")# 关闭环境env.close()return {'episodes': range(len(rewards)), 'rewards': rewards}# 测试函数
def test(arg_dict, env, agent):startTime = time.time()print("开始测试智能体......")print(f"环境名: {arg_dict['env_name']}, 算法名: {arg_dict['algo_name']}, Device: {arg_dict['device']}")rewards = []steps = []for i_ep in range(arg_dict['test_eps']):ep_reward = 0ep_step = 0state = env.reset()for _ in range(arg_dict['ep_max_steps']):# 画图if arg_dict['test_render']:env.render()ep_step += 1action = agent.predict_action(state)next_state, reward, done, _ = env.step(action)state = next_stateep_reward += rewardif done:breaksteps.append(ep_step)rewards.append(ep_reward)print(f"Episode: {i_ep + 1}/{arg_dict['test_eps']}，Reward: {ep_reward:.2f}")print("测试结束 , 用时: " + str(time.time() - startTime) + " s")env.close()return {'episodes': range(len(rewards)), 'rewards': rewards}# 创建环境和智能体
def create_env_agent(arg_dict):# 创建环境env = gym.make(arg_dict['env_name'])# 设置随机种子all_seed(env, seed=arg_dict["seed"])# 获取状态数try:n_states = env.observation_space.nexcept AttributeError:n_states = env.observation_space.shape[0]# 获取动作数n_actions = env.action_space.nprint(f"状态数: {n_states}, 动作数: {n_actions}")# 将状态数和动作数加入算法参数字典arg_dict.update({"n_states": n_states, "n_actions": n_actions})# 实例化智能体对象# Q网络模型model = MLP(n_states, n_actions, hidden_dim=arg_dict["hidden_dim"])# 回放缓存区对象memory = ReplayBuffer(arg_dict["memory_capacity"])# 智能体agent = DQN(model, memory, arg_dict)# 返回环境，智能体return env, agentif __name__ == '__main__':# 防止报错 OMP: Error #15: Initializing libiomp5md.dll, but found libiomp5md.dll already initialized.os.environ["KMP_DUPLICATE_LIB_OK"] = "TRUE"# 获取当前路径curr_path = os.path.dirname(os.path.abspath(__file__))# 获取当前时间curr_time = datetime.datetime.now().strftime("%Y_%m_%d-%H_%M_%S")# 相关参数设置parser = argparse.ArgumentParser(description="hyper parameters")parser.add_argument('--algo_name', default='DQN', type=str, help="name of algorithm")parser.add_argument('--env_name', default='CartPole-v0', type=str, help="name of environment")parser.add_argument('--train_eps', default=200, type=int, help="episodes of training")parser.add_argument('--test_eps', default=20, type=int, help="episodes of testing")parser.add_argument('--ep_max_steps', default=100000, type=int,help="steps per episode, much larger value can simulate infinite steps")parser.add_argument('--gamma', default=0.95, type=float, help="discounted factor")parser.add_argument('--epsilon_start', default=0.95, type=float, help="initial value of epsilon")parser.add_argument('--epsilon_end', default=0.01, type=float, help="final value of epsilon")parser.add_argument('--epsilon_decay', default=500, type=int,help="decay rate of epsilon, the higher value, the slower decay")parser.add_argument('--lr', default=0.0001, type=float, help="learning rate")parser.add_argument('--memory_capacity', default=100000, type=int, help="memory capacity")parser.add_argument('--batch_size', default=64, type=int)parser.add_argument('--target_update', default=4, type=int)parser.add_argument('--hidden_dim', default=256, type=int)parser.add_argument('--device', default='cpu', type=str, help="cpu or cuda")parser.add_argument('--seed', default=520, type=int, help="seed")parser.add_argument('--show_fig', default=False, type=bool, help="if show figure or not")parser.add_argument('--save_fig', default=True, type=bool, help="if save figure or not")parser.add_argument('--train_render', default=False, type=bool,help="Whether to render the environment during training")parser.add_argument('--test_render', default=True, type=bool,help="Whether to render the environment during testing")args = parser.parse_args()default_args = {'result_path': f"{curr_path}/outputs/{args.env_name}/{curr_time}/results/",'model_path': f"{curr_path}/outputs/{args.env_name}/{curr_time}/models/",}# 将参数转化为字典 type(dict)arg_dict = {**vars(args), **default_args}print("算法参数字典:", arg_dict)# 创建环境和智能体env, agent = create_env_agent(arg_dict)# 传入算法参数、环境、智能体，然后开始训练res_dic = train(arg_dict, env, agent)print("算法返回结果字典:", res_dic)# 保存相关信息agent.save_model(path=arg_dict['model_path'])save_args(arg_dict, path=arg_dict['result_path'])save_results(res_dic, tag='train', path=arg_dict['result_path'])plot_rewards(res_dic['rewards'], arg_dict, path=arg_dict['result_path'], tag="train")# =================================================================================================# 创建新环境和智能体用来测试print("=" * 300)env, agent = create_env_agent(arg_dict)# 加载已保存的智能体agent.load_model(path=arg_dict['model_path'])res_dic = test(arg_dict, env, agent)save_results(res_dic, tag='test', path=arg_dict['result_path'])plot_rewards(res_dic['rewards'], arg_dict, path=arg_dict['result_path'], tag="test")

4.3 运行结果展示

由于有些输出太长，下面仅展示部分输出

状态数: 4, 动作数: 2
环境名: CartPole-v0, 算法名: DQN, Device: cpu
开始训练智能体......
Begin to update!
Episode: 10/200, Reward: 16.00: Epislon: 0.649
Episode: 20/200, Reward: 12.00: Epislon: 0.473
Episode: 30/200, Reward: 10.00: Epislon: 0.358
Episode: 40/200, Reward: 17.00: Epislon: 0.272
Episode: 50/200, Reward: 18.00: Epislon: 0.212
Episode: 60/200, Reward: 139.00: Epislon: 0.049
Episode: 70/200, Reward: 200.00: Epislon: 0.011
Episode: 80/200, Reward: 200.00: Epislon: 0.010
Episode: 90/200, Reward: 200.00: Epislon: 0.010
Episode: 100/200, Reward: 200.00: Epislon: 0.010
Episode: 110/200, Reward: 200.00: Epislon: 0.010
Episode: 120/200, Reward: 200.00: Epislon: 0.010
Episode: 130/200, Reward: 169.00: Epislon: 0.010
Episode: 140/200, Reward: 200.00: Epislon: 0.010
Episode: 150/200, Reward: 179.00: Epislon: 0.010
Episode: 160/200, Reward: 200.00: Epislon: 0.010
Episode: 170/200, Reward: 170.00: Epislon: 0.010
Episode: 180/200, Reward: 200.00: Epislon: 0.010
Episode: 190/200, Reward: 200.00: Epislon: 0.010
Episode: 200/200, Reward: 165.00: Epislon: 0.010
训练结束 , 用时: 100.28473830223083 s
============================================================================================================================================================================================================================================================================================================
状态数: 4, 动作数: 2
开始测试智能体......
环境名: CartPole-v0, 算法名: DQN, Device: cpu
Episode: 1/20，Reward: 200.00
Episode: 2/20，Reward: 200.00
Episode: 3/20，Reward: 200.00
Episode: 4/20，Reward: 200.00
Episode: 5/20，Reward: 200.00
Episode: 6/20，Reward: 200.00
Episode: 7/20，Reward: 200.00
Episode: 8/20，Reward: 200.00
Episode: 9/20，Reward: 200.00
Episode: 10/20，Reward: 200.00
Episode: 11/20，Reward: 200.00
Episode: 12/20，Reward: 198.00
Episode: 13/20，Reward: 200.00
Episode: 14/20，Reward: 200.00
Episode: 15/20，Reward: 200.00
Episode: 16/20，Reward: 200.00
Episode: 17/20，Reward: 200.00
Episode: 18/20，Reward: 179.00
Episode: 19/20，Reward: 200.00
Episode: 20/20，Reward: 200.00
测试结束 , 用时: 30.37125039100647 s