简单DDPG（参考知乎花半楼）

import gym
import math
import random
import numpy as np
import matplotlib.pyplot as plt
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
import os
'''迁移到了GPU上进行训练'''

class Actor(nn.Module):
    def __init__(self, input_size, hidden_size, output_size):
        super(Actor, self).__init__()
        self.linear1 = nn.Linear(input_size, hidden_size)
        self.linear2 = nn.Linear(hidden_size, hidden_size)
        self.linear3 = nn.Linear(hidden_size, output_size)

    def forward(self, s):
        x = F.relu(self.linear1(s))
        x = F.relu(self.linear2(x))
        x = torch.tanh(self.linear3(x))

        return x


class Critic(nn.Module):
    def __init__(self, input_size, hidden_size, output_size):
        super().__init__()
        self.linear1 = nn.Linear(input_size, hidden_size)
        self.linear2 = nn.Linear(hidden_size, hidden_size)
        self.linear3 = nn.Linear(hidden_size, output_size)

    def forward(self, s, a):
        x = torch.cat([s, a], 1)
        x = F.relu(self.linear1(x))
        x = F.relu(self.linear2(x))
        x = self.linear3(x)

        return x


class Agent(object):
    def __init__(self, **kwargs):
        for key, value in kwargs.items():
            setattr(self, key, value)

        s_dim = self.env.observation_space.shape[0]
        a_dim = self.env.action_space.shape[0]
        fileList = os.listdir('nets/')
        if "actor.pkl" in fileList :
            print("Trained nets found!")
            self.actor = torch.load('nets/actor.pkl')
            self.actor_target = torch.load('nets/actor_target.pkl')
            self.critic = torch.load('nets/critic.pkl')
            self.critic_target = torch.load('nets/critic_target.pkl')
        else:
            print("Trained nets not found!")
            self.actor = Actor(s_dim, 256, a_dim).cuda()
            self.actor_target = Actor(s_dim, 256, a_dim).cuda()
            self.critic = Critic(s_dim + a_dim, 256, 1).cuda()   # 此处修改了critic的输出维度恒为1
            self.critic_target = Critic(s_dim + a_dim, 256, 1).cuda()  # 此处修改了critic的输出维度恒为1
        self.actor_optim = optim.Adam(self.actor.parameters(), lr=self.actor_lr)
        self.critic_optim = optim.Adam(self.critic.parameters(), lr=self.critic_lr)
        self.buffer = []

        self.actor_target.load_state_dict(self.actor.state_dict())
        self.critic_target.load_state_dict(self.critic.state_dict())

    def act(self, s0):
        s0 = torch.tensor(s0, dtype=torch.float).cuda().unsqueeze(0).cuda()
        a0 = self.actor(s0).squeeze(0).detach().cpu().numpy()
        return a0

    def put(self, *transition):
        if len(self.buffer) == self.capacity:
            self.buffer.pop(0)
        self.buffer.append(transition)

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

        samples = random.sample(self.buffer, self.batch_size)

        s0, a0, r1, s1 = zip(*samples)

        s0 = torch.tensor(s0, dtype=torch.float).cuda()
        a0 = torch.tensor(a0, dtype=torch.float).cuda()
        r1 = torch.tensor(r1, dtype=torch.float).view(self.batch_size, -1).cuda()
        s1 = torch.tensor(s1, dtype=torch.float).cuda()

        def critic_learn():
            a1 = self.actor_target(s1).detach()
            y_true = r1 + self.gamma * self.critic_target(s1, a1).detach()

            y_pred = self.critic(s0, a0)

            loss_fn = nn.MSELoss()
            loss = loss_fn(y_pred, y_true)
            self.critic_optim.zero_grad()
            loss.backward()
            self.critic_optim.step()

        def actor_learn():
            loss = -torch.mean(self.critic(s0, self.actor(s0)))
            self.actor_optim.zero_grad()
            loss.backward()
            self.actor_optim.step()

        def soft_update(net_target, net, tau):
            for target_param, param in zip(net_target.parameters(), net.parameters()):
                target_param.data.copy_(target_param.data * (1.0 - tau) + param.data * tau)

        critic_learn()
        actor_learn()
        soft_update(self.critic_target, self.critic, self.tau)
        soft_update(self.actor_target, self.actor, self.tau)

    def save(self):
        torch.save(self.actor, 'nets/actor.pkl')
        torch.save(self.actor_target, 'nets/actor_target.pkl')
        torch.save(self.critic, 'nets/critic.pkl')
        torch.save(self.critic_target, 'nets/critic_target.pkl')

env = gym.make('Pendulum-v1')
env.reset()
env.render()

params = {
    'env': env,
    'gamma': 0.99,
    'actor_lr': 0.001,
    'critic_lr': 0.0013,
    'tau': 0.02,
    'capacity': 20000,
    'batch_size': 32,
}
EPOCH_NUM = 5
agent = Agent(**params)
FLAG = False
rewardList = []
for episode in range(EPOCH_NUM):
    s0 = env.reset()
    episode_reward = 0
    if(episode%10 == 0):
        flag = True
    else:
        flag = False

    for step in range(500):
        if(flag):
            env.render()
        a0 = agent.act(s0)
        s1, r1, done, _ = env.step(a0)
        agent.put(s0, a0, r1, s1)

        episode_reward += r1
        s0 = s1

        agent.learn()

    print(episode, ': ', episode_reward)
    rewardList.append(episode_reward)
pltX = [i for i in range(EPOCH_NUM)]
plt.plot(pltX,rewardList)
plt.show()
agent.save()

更改了原代码中Critic输出数据维度不对（不是1）的问题，并且将代码迁移到了GPU训练，增加了保存和加载功能

利用gym的 Pendulum-v1训练750个episode的reward结果

原文链接

强化学习-基于pytorch的DDPG实现 - 知乎 (zhihu.com)https://zhuanlan.zhihu.com/p/65931777