从DQN到Double DQN和Dueling DQN——pytorch实操

写这个文章的动机是一直没有人讲明白三种DQN之间的关系，要不过于学术，要不过于工业界。本文试图两方面结合，说说改了什么，有什么好处，效果如何。当然，也是干货满满，直接上代码。

听说点进蝈仔帖子的都喜欢点赞加关注~~
鸣谢：
李宏毅教授 http://speech.ee.ntu.edu.tw/~tlkagk/index.html
深度强化学习中文版一书
知乎专栏：
https://zhuanlan.zhihu.com/p/131662954

DQN

首先附上难得的中文算法表

原理这边简单带过：
先是Q-learning

之后使用网络价值函数逼近：

说白了就是个网络拟合

直接代码附上

这是最经典的小摆锤

# -*- coding: utf-8 -*-
# import the necessary packages
import torch
import torch.nn as nn
from torch.autograd import Variable
import torch.nn.functional as F
import numpy as np
import gym

# 1. Define some Hyper Parameters
BATCH_SIZE = 32     # batch size of sampling process from buffer
LR = 0.01           # learning rate
EPSILON = 0.9       # epsilon used for epsilon greedy approach
GAMMA = 0.9         # discount factor
TARGET_NETWORK_REPLACE_FREQ = 100       # How frequently target netowrk updates
MEMORY_CAPACITY = 2000                  # The capacity of experience replay buffer

env = gym.make("CartPole-v0") # Use cartpole game as environment
env = env.unwrapped
N_ACTIONS = env.action_space.n  # 2 actions
N_STATES = env.observation_space.shape[0] # 4 states
ENV_A_SHAPE = 0 if isinstance(env.action_space.sample(), int) else env.action_space.sample().shape     # to confirm the shape

# 2. Define the network used in both target net and the net for training
class Net(nn.Module):
    def __init__(self):
        # Define the network structure, a very simple fully connected network
        super(Net, self).__init__()
        # Define the structure of fully connected network
        self.fc1 = nn.Linear(N_STATES, 10)  # layer 1
        self.fc1.weight.data.normal_(0, 0.1) # in-place initilization of weights of fc1
        self.out = nn.Linear(10, N_ACTIONS) # layer 2
        self.out.weight.data.normal_(0, 0.1) # in-place initilization of weights of fc2
        
        
    def forward(self, x):
        # Define how the input data pass inside the network
        x = self.fc1(x)
        x = F.relu(x)
        actions_value = self.out(x)
        return actions_value
        
# 3. Define the DQN network and its corresponding methods
class DQN(object):
    def __init__(self):
        # -----------Define 2 networks (target and training)------#
        self.eval_net, self.target_net = Net(), Net()
        # Define counter, memory size and loss function
        self.learn_step_counter = 0 # count the steps of learning process
        self.memory_counter = 0 # counter used for experience replay buffer
        
        # ----Define the memory (or the buffer), allocate some space to it. The number 
        # of columns depends on 4 elements, s, a, r, s_, the total is N_STATES*2 + 2---#
        self.memory = np.zeros((MEMORY_CAPACITY, N_STATES * 2 + 2)) 
        
        #------- Define the optimizer------#
        self.optimizer = torch.optim.Adam(self.eval_net.parameters(), lr=LR)
        
        # ------Define the loss function-----#
        self.loss_func = nn.MSELoss()
        
    def  choose_action(self, x):
        # This function is used to make decision based upon epsilon greedy
        
        x = torch.unsqueeze(torch.FloatTensor(x), 0) # add 1 dimension to input state x
        # input only one sample
        if np.random.uniform() < EPSILON:   # greedy
            # use epsilon-greedy approach to take action
            actions_value = self.eval_net.forward(x)
            #print(torch.max(actions_value, 1)) 
            # torch.max() returns a tensor composed of max value along the axis=dim and corresponding index
            # what we need is the index in this function, representing the action of cart.
            action = torch.max(actions_value, 1)[1].data.numpy()
            action = action[0] if ENV_A_SHAPE == 0 else action.reshape(ENV_A_SHAPE)  # return the argmax index
        else:   # random
            action = np.random.randint(0, N_ACTIONS)
            action = action if ENV_A_SHAPE == 0 else action.reshape(ENV_A_SHAPE)
        return action
    
        
    def store_transition(self, s, a, r, s_):
        # This function acts as experience replay buffer        
        transition = np.hstack((s, [a, r], s_)) # horizontally stack these vectors
        # if the capacity is full, then use index to replace the old memory with new one
        index = self.memory_counter % MEMORY_CAPACITY
        self.memory[index, :] = transition
        self.memory_counter += 1
        
    
    def learn(self):
        # Define how the whole DQN works including sampling batch of experiences,
        # when and how to update parameters of target network, and how to implement
        # backward propagation.
        
        # update the target network every fixed steps
        if self.learn_step_counter % TARGET_NETWORK_REPLACE_FREQ == 0:
            # Assign the parameters of eval_net to target_net
            self.target_net.load_state_dict(self.eval_net.state_dict())
        self.learn_step_counter += 1
        
        # Determine the index of Sampled batch from buffer
        sample_index = np.random.choice(MEMORY_CAPACITY, BATCH_SIZE) # randomly select some data from buffer
        # extract experiences of batch size from buffer.
        b_memory = self.memory[sample_index, :]
        # extract vectors or matrices s,a,r,s_ from batch memory and convert these to torch Variables
        # that are convenient to back propagation
        b_s = Variable(torch.FloatTensor(b_memory[:, :N_STATES]))
        # convert long int type to tensor
        b_a = Variable(torch.LongTensor(b_memory[:, N_STATES:N_STATES+1].astype(int)))
        b_r = Variable(torch.FloatTensor(b_memory[:, N_STATES+1:N_STATES+2]))
        b_s_ = Variable(torch.FloatTensor(b_memory[:, -N_STATES:]))
        
        # calculate the Q value of state-action pair
        q_eval = self.eval_net(b_s).gather(1, b_a) # (batch_size, 1)
        #print(q_eval)
        # calculate the q value of next state
        q_next = self.target_net(b_s_).detach() # detach from computational graph, don't back propagate
        # select the maximum q value
        #print(q_next)
        # q_next.max(1) returns the max value along the axis=1 and its corresponding index
        q_target = b_r + GAMMA * q_next.max(1)[0].view(BATCH_SIZE, 1) # (batch_size, 1)
        loss = self.loss_func(q_eval, q_target)
        
        self.optimizer.zero_grad() # reset the gradient to zero
        loss.backward()
        self.optimizer.step() # execute back propagation for one step
 
'''
--------------Procedures of DQN Algorithm------------------
'''
# create the object of DQN class
dqn = DQN()

# Start training
print("\nCollecting experience...")
for i_episode in range(400):
    # play 400 episodes of cartpole game
    s = env.reset()
    ep_r = 0
    while True:
        env.render()
        # take action based on the current state
        a = dqn.choose_action(s)
        # obtain the reward and next state and some other information
        s_, r, done, info = env.step(a)
        
        # modify the reward based on the environment state
        x, x_dot, theta, theta_dot = s_
        r1 = (env.x_threshold - abs(x)) / env.x_threshold - 0.8
        r2 = (env.theta_threshold_radians - abs(theta)) / env.theta_threshold_radians - 0.5
        r = r1 + r2
        
        # store the transitions of states
        dqn.store_transition(s, a, r, s_)
        
        ep_r += r
        # if the experience repaly buffer is filled, DQN begins to learn or update
        # its parameters.       
        if dqn.memory_counter > MEMORY_CAPACITY:
            dqn.learn()
            if done:
                print('Ep: ', i_episode, ' |', 'Ep_r: ', round(ep_r, 2))
        
        if done:
            # if game is over, then skip the while loop.
            break
        # use next state to update the current state. 
        s = s_  

为什么会有改进

DQN的实践过程中会出现一些问题，比如高估了动作值（overestimation）。

Double DQN

Double DQN其实就是Double Q learning在DQN上的拓展，上面Q和Q2两套Q值，分别对应DQN的policy network（更新的快）和target network（每隔一段时间与policy network同步）。
换言之：
Double DQN(DDQN)是DQN的一种改进。在DDQN之前，基本所有的目标Q值都是通过贪婪法得到的，而这往往会造成过度估计(overestimations)的问题。DDQN将目标Q值的最大动作分解成动作选择和动作评估两步，有效解决了这个问题。
Double Q learning error如下：

Double DQN 相比于 DQN 的效果提升情况如图

代码附上

实现起来也很简单，只需要在DQN计算TD error的时候稍作改动：

def compute_td_loss(self, states, actions, rewards, next_states, is_done, gamma=0.99):
    """ Compute td loss using torch operations only. Use the formula above. """
    actions = torch.tensor(actions).long()    # shape: [batch_size]
    rewards = torch.tensor(rewards, dtype =torch.float)  # shape: [batch_size]
    is_done = torch.tensor(done).bool()  # shape: [batch_size]

    if self.USE_CUDA:
        actions = actions.cuda()
        rewards = rewards.cuda()
        is_done = is_done.cuda()

    # get q-values for all actions in current states
    predicted_qvalues = self.DQN(states)

    # select q-values for chosen actions
    predicted_qvalues_for_actions = predicted_qvalues[
      range(states.shape[0]), actions
    ]

    # compute q-values for all actions in next states 
    ## Where DDQN is different from DQN
    predicted_next_qvalues_current = self.DQN(next_states)
    predicted_next_qvalues_target = self.DQN_target(next_states)
    # compute V*(next_states) using predicted next q-values
    next_state_values =  predicted_next_qvalues_target.gather(1, torch.max(predicted_next_qvalues_current, 1)[1].unsqueeze(1)).squeeze(1)

    # compute "target q-values" for loss - it's what's inside square parentheses in the above formula.
    target_qvalues_for_actions = rewards + gamma *next_state_values # YOUR CODE

    # at the last state we shall use simplified formula: Q(s,a) = r(s,a) since s' doesn't exist
    target_qvalues_for_actions = torch.where(
        is_done, rewards, target_qvalues_for_actions)

    # mean squared error loss to minimize
    #loss = torch.mean((predicted_qvalues_for_actions -
    #                   target_qvalues_for_actions.detach()) ** 2)
    loss = F.smooth_l1_loss(predicted_qvalues_for_actions, target_qvalues_for_actions.detach())

    return loss

Dueling DQN

这个改的是网络结构，上图好理解：


效果提升如图所示：

话不多说直接给代码

Dueling DQN在DQN上的修改也很小，只是多出一支fully connected layer来估计V(s)。代码如下：

class Dueling_DQN(nn.Module):
    def __init__(self, input_shape, num_outputs):
        super(Dueling_DQN, self).__init__()
        
        self.input_shape = input_shape
        self.num_actions = num_outputs
        
        self.features = nn.Sequential(
            nn.Conv2d(input_shape[0], 32, kernel_size=8, stride=4),
            nn.ReLU(),
            nn.Conv2d(32, 64, kernel_size=4, stride=2),
            nn.ReLU(),
            nn.Conv2d(64, 64, kernel_size=3, stride=1),
            nn.ReLU()
        )
        
        self.advantage = nn.Sequential(
            nn.Linear(self.feature_size(), 512),
            nn.ReLU(),
            nn.Linear(512, num_outputs)
        )
        
        self.value = nn.Sequential(
            nn.Linear(self.feature_size(), 512),
            nn.ReLU(),
            nn.Linear(512, 1)
        )
        
    def forward(self, x):
        x = self.features(x)
        x = x.view(x.size(0), -1)
        advantage = self.advantage(x)
        value     = self.value(x)
        return value + advantage  - advantage.mean()
    
    def feature_size(self):
        return self.features(autograd.Variable(torch.zeros(1, *self.input_shape))).view(1, -1).size(1)

改进究竟管用与否？

下面是DDQN（蓝色），Dueling DQN（粉红）和DQN（橘黄）在Pong上面得表现。可以看到DDQN > Dueling DQN > DQN，其中DDQN大约比DQN得收敛快10%。