强化学习经典算法笔记(十三)：深度确定性策略梯度算法DDPG的pytorch实现

最新推荐文章于 2024-09-13 08:06:13 发布

hhy_csdn

最新推荐文章于 2024-09-13 08:06:13 发布

阅读量3.1k

点赞数 4

分类专栏：强化学习文章标签：深度学习强化学习算法强化学习 pytorch 深度学习机器学习

本文链接：https://blog.csdn.net/hhy_csdn/article/details/107060245

版权

强化学习专栏收录该内容

38 篇文章 161 订阅

订阅专栏

强化学习经典算法笔记(十三)：深度确定性策略梯度算法DDPG的pytorch实现

一、DDPG算法的要点

DDPG适用于连续动作空间的控制任务
DDPG解决了DQN难以对连续动作估计Q值的问题
确定性策略是指：在某个状态 $s_t$ 所采取的动作是确定的。由此带来了Bellman方程的改变。由
$Q^{\pi}(s_t,a_t)=E_{s_{t+1}\sim E,a_t \sim \pi}[r(s_t,a_t)+\gamma E_{\pi}[Q^{\pi}(s_{t+1},a_{t+1})]]$
变成了
$Q^{\mu}(s_t,a_t)=E_{s_{t+1}\sim E}[r(s_t,a_t)+\gamma Q^{\mu}(s_{t+1},\mu(s_{t+1})]$
区别在于确定性动作 $\mu(s_t)$ 取代了从随机性策略中采样的动作 $a_t \sim \pi(a|s_t)$ ，因此中括号内部对策略求期望的操作也省去了。只需要对环境求期望即可。
也就是说动作-状态值函数Q只和环境有关系，也就意味着外面可以使用off-policy来更新值函数（比如使用Q-learning方法等）。
使用Actor-critic框架，Actor输入状态，输出确定性动作，Critic输入状态和动作，输出Q值。
借鉴DQN，使用了Memory Buffer和Target Networks，对Critic网络做Off-policy的更新。
使用 soft target update，缓慢更新目标网络。
$\theta' \leftarrow \tau \theta + (1-\tau)\theta',\; \tau \ll 1$
使用OU噪声，一种时序噪声，为确定性动作提供exploration的能力。
使用batch normalization。

图片来自DDPG论文笔记。

二、DDPG的Pytorch实现

import torch
import torch.nn as nn
import torch.nn.functional as F
import numpy as np
import gym
import time


#####################  hyper parameters  ####################

MAX_EPISODES = 200               # 最大训练代数
MAX_EP_STEPS = 200               # episode最大持续帧数

RENDER = False
ENV_NAME = 'Pendulum-v0'         # 游戏名称
SEED = 123                       # 随机数种子

DDPG算法主体的实现。由于动作向量的取值范围是对称的，所以输入只有一个a_bound。

###############################  DDPG  ####################################
class DDPG(object):
    def __init__(self, a_dim, s_dim, a_bound,):
        self.a_dim = a_dim
        self.s_dim = s_dim
        self.a_bound = a_bound
        self.pointer = 0                                                                         # exp buffer指针
        self.lr_a = 0.001                                                                        # learning rate for actor
        self.lr_c = 0.002                                                                        # learning rate for critic
        self.gamma = 0.9                                                                         # reward discount
        self.tau = 0.01                                                                          # 软更新比例
        self.memory_capacity = 10000
        self.batch_size = 32
        self.memory = np.zeros((self.memory_capacity, s_dim * 2 + a_dim + 1), dtype=np.float32)

        
        class ANet(nn.Module):                               # 定义动作网络
            def __init__(self, s_dim, a_dim, a_bound):
                super(ANet,self).__init__()
                self.a_bound = a_bound
                self.fc1 = nn.Linear(s_dim,30)
                self.fc1.weight.data.normal_(0,0.1)          # initialization
                
                self.out = nn.Linear(30,a_dim)
                self.out.weight.data.normal_(0,0.1)          # initialization
            def forward(self,x):
                x = self.fc1(x)
                x = F.relu(x)
                x = self.out(x)
                x = F.tanh(x)
                actions_value = x * self.a_bound.item()
                return actions_value

        class CNet(nn.Module):                               # 定义价值网络
            def __init__(self,s_dim,a_dim):
                super(CNet,self).__init__()
                self.fcs = nn.Linear(s_dim,30)
                self.fcs.weight.data.normal_(0,0.1)          # initialization
                
                self.fca = nn.Linear(a_dim,30)
                self.fca.weight.data.normal_(0,0.1)          # initialization
                
                self.out = nn.Linear(30,1)
                self.out.weight.data.normal_(0, 0.1)         # initialization
            def forward(self,s,a):
                x = self.fcs(s)                              # 输入状态
                y = self.fca(a)                              # 输入动作
                net = F.relu(x+y)
                actions_value = self.out(net)                # 给出V(s,a)
                return actions_value

        self.Actor_eval = ANet(s_dim, a_dim, a_bound)        # 主网络
        self.Actor_target = ANet(s_dim, a_dim, a_bound)      # 目标网络
        self.Critic_eval = CNet(s_dim, a_dim)                # 主网络
        self.Critic_target = CNet(s_dim, a_dim)              # 当前网络
        self.ctrain = torch.optim.Adam(self.Critic_eval.parameters(),lr = self.lr_c) # critic的优化器
        self.atrain = torch.optim.Adam(self.Actor_eval.parameters(),lr = self.lr_a)  # actor的优化器
        self.loss_td = nn.MSELoss()                          # 损失函数采用均方误差

    def choose_action(self, s):
        s = torch.unsqueeze(torch.FloatTensor(s), 0)
        return self.Actor_eval(s)[0].detach()                # detach()不需要计算梯度

    def learn(self):

        for x in self.Actor_target.state_dict().keys():
            eval('self.Actor_target.' + x + '.data.mul_((1 - self.tau))')  
            eval('self.Actor_target.' + x + '.data.add_(self.tau * self.Actor_eval.' + x + '.data)')
        for x in self.Critic_target.state_dict().keys():
            eval('self.Critic_target.' + x + '.data.mul_((1- self.tau))')
            eval('self.Critic_target.' + x + '.data.add_(self.tau * self.Critic_eval.' + x + '.data)')

        # soft target replacement

        indices = np.random.choice(self.memory_capacity, size = self.batch_size)  # 随机采样的index
        bt = self.memory[indices, :]                                              # 采样batch_size个sample
        bs = torch.FloatTensor(bt[:, :self.s_dim])                                # state
        ba = torch.FloatTensor(bt[:, self.s_dim: self.s_dim + self.a_dim])        # action
        br = torch.FloatTensor(bt[:, -self.s_dim - 1: -self.s_dim])               # reward
        bs_ = torch.FloatTensor(bt[:, -self.s_dim:])                              # next state

        a = self.Actor_eval(bs)
        q = self.Critic_eval(bs,a)  # loss=-q=-ce(s,ae(s))更新ae   ae(s)=a   ae(s_)=a_
        # 如果 a是一个正确的行为的话，那么它的Q应该更贴近0
        loss_a = -torch.mean(q) 
        # print(q)
        # print(loss_a)
        self.atrain.zero_grad()
        loss_a.backward()
        self.atrain.step()

        a_ = self.Actor_target(bs_)      # 这个网络不及时更新参数, 用于预测 Critic 的 Q_target 中的 action
        q_ = self.Critic_target(bs_,a_)  # 这个网络不及时更新参数, 用于给出 Actor 更新参数时的 Gradient ascent 强度
        q_target = br + self.gamma * q_  # q_target = 负的
        #print(q_target)
        q_v = self.Critic_eval(bs,ba)
        #print(q_v)
        td_error = self.loss_td(q_target,q_v)
        # td_error = R + self.gamma * ct（bs_,at(bs_)）-ce(s,ba) 更新ce ,但这个ae(s)是记忆中的ba，让ce得出的Q靠近Q_target,让评价更准确
        #print(td_error)
        self.ctrain.zero_grad()
        td_error.backward()
        self.ctrain.step()

    def store_transition(self, s, a, r, s_):
        transition = np.hstack((s, a, [r], s_))
        index = self.pointer % self.memory_capacity     # replace the old memory with new memory
        self.memory[index, :] = transition
        self.pointer += 1                               # 指示sample位置的指针+1

训练代码。

###############################  training  ####################################
env = gym.make(ENV_NAME)
env = env.unwrapped
env.seed(SEED)                                          # 设置Gym的随机数种子
torch.manual_seed(SEED)                                 # 设置pytorch的随机数种子

s_dim = env.observation_space.shape[0]                  # 状态空间
a_dim = env.action_space.shape[0]                       # 动作空间
a_bound = env.action_space.high                         # 动作取值区间,对称区间，故只取上界
ddpg = DDPG(a_dim, s_dim, a_bound)

var = 3                                                 # 动作服从的高斯分布的方差，控制探索程度
t1 = time.time()                                        # 开始时间
for i in range(MAX_EPISODES):
    s = env.reset()
    ep_reward = 0
    for j in range(MAX_EP_STEPS):
        if RENDER:
            env.render()

        # Add exploration noise
        a = ddpg.choose_action(s)
        a = np.clip(np.random.normal(a, var), -2, 2)    # add randomness to action selection for exploration
        s_, r, done, info = env.step(a)

        ddpg.store_transition(s, a, r / 10, s_)         # 为什么要对reward归一化

        if ddpg.pointer > ddpg.memory_capacity:              # 经验池已满
            var *= .9995                                # 学习阶段逐渐降低动作随机性decay the action randomness
            ddpg.learn()                                # 开始学习

        s = s_
        ep_reward += r
        if j == MAX_EP_STEPS-1:
            print('Episode:', i, ' Reward: %i' % int(ep_reward), 'Explore: %.2f' % var, )
            # if ep_reward > -300:
                # RENDER = True
            break
print('Running time: ', time.time() - t1)