文章目錄

• DQN
• • 直接代碼附上
• 為什么會有改進
• • Double DQN
  • • 代碼附上
  • Dueling DQN
  • • 話不多說直接給代碼
• 改進究竟管用與否？

寫這個文章的動機是一直沒有人講明白三種DQN之間的關系，要不過于學術，要不過于工業(yè)界。本文試圖兩方面結(jié)合，說說改了什么，有什么好處，效果如何。當然，也是干貨滿滿，直接上代碼。

聽說點進蟈仔帖子的都喜歡點贊加關注~~
鳴謝：
李宏毅教授 http://speech.ee.ntu.edu.tw/~tlkagk/index.html
深度強化學習中文版一書
知乎專欄：
https://zhuanlan.zhihu.com/p/131662954

DQN

首先附上難得的中文算法表

原理這邊簡單帶過：
先是Q-learning

之后使用網(wǎng)絡價值函數(shù)逼近：

說白了就是個網(wǎng)絡擬合

直接代碼附上

這是最經(jīng)典的小擺錘

# -*- 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 bufferenv = 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 networksuper(Net, self).__init__()# Define the structure of fully connected networkself.fc1 = nn.Linear(N_STATES, 10) # layer 1self.fc1.weight.data.normal_(0, 0.1) # in-place initilization of weights of fc1self.out = nn.Linear(10, N_ACTIONS) # layer 2self.out.weight.data.normal_(0, 0.1) # in-place initilization of weights of fc2def forward(self, x):# Define how the input data pass inside the networkx = 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 functionself.learn_step_counter = 0 # count the steps of learning processself.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 greedyx = torch.unsqueeze(torch.FloatTensor(x), 0) # add 1 dimension to input state x# input only one sampleif np.random.uniform() < EPSILON: # greedy# use epsilon-greedy approach to take actionactions_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 indexelse: # randomaction = np.random.randint(0, N_ACTIONS)action = action if ENV_A_SHAPE == 0 else action.reshape(ENV_A_SHAPE)return actiondef 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 oneindex = self.memory_counter % MEMORY_CAPACITYself.memory[index, :] = transitionself.memory_counter += 1def 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 stepsif self.learn_step_counter % TARGET_NETWORK_REPLACE_FREQ == 0:# Assign the parameters of eval_net to target_netself.target_net.load_state_dict(self.eval_net.state_dict())self.learn_step_counter += 1# Determine the index of Sampled batch from buffersample_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 propagationb_s = Variable(torch.FloatTensor(b_memory[:, :N_STATES]))# convert long int type to tensorb_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 pairq_eval = self.eval_net(b_s).gather(1, b_a) # (batch_size, 1)#print(q_eval)# calculate the q value of next stateq_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 indexq_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 zeroloss.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 games = env.reset()ep_r = 0while True:env.render()# take action based on the current statea = dqn.choose_action(s)# obtain the reward and next state and some other informations_, r, done, info = env.step(a)# modify the reward based on the environment statex, x_dot, theta, theta_dot = s_r1 = (env.x_threshold - abs(x)) / env.x_threshold - 0.8r2 = (env.theta_threshold_radians - abs(theta)) / env.theta_threshold_radians - 0.5r = r1 + r2# store the transitions of statesdqn.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_

為什么會有改進

DQN的實踐過程中會出現(xiàn)一些問題，比如高估了動作值（overestimation）。

Double DQN

Double DQN其實就是Double Q learning在DQN上的拓展，上面Q和Q2兩套Q值，分別對應DQN的policy network（更新的快）和target network（每隔一段時間與policy network同步）。
換言之：
Double DQN(DDQN)是DQN的一種改進。在DDQN之前，基本所有的目標Q值都是通過貪婪法得到的，而這往往會造成過度估計(overestimations)的問題。DDQN將目標Q值的最大動作分解成動作選擇和動作評估兩步，有效解決了這個問題。
Double Q learning error如下：

Double DQN 相比于 DQN 的效果提升情況如圖

代碼附上

實現(xiàn)起來也很簡單，只需要在DQN計算TD error的時候稍作改動：

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 statespredicted_qvalues = self.DQN(states)# select q-values for chosen actionspredicted_qvalues_for_actions = predicted_qvalues[range(states.shape[0]), actions]# compute q-values for all actions in next states ## Where DDQN is different from DQNpredicted_next_qvalues_current = self.DQN(next_states)predicted_next_qvalues_target = self.DQN_target(next_states)# compute V*(next_states) using predicted next q-valuesnext_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 existtarget_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

這個改的是網(wǎng)絡結(jié)構(gòu)，上圖好理解：


效果提升如圖所示：

話不多說直接給代碼

Dueling DQN在DQN上的修改也很小，只是多出一支fully connected layer來估計V(s)。代碼如下：

class Dueling_DQN(nn.Module):def __init__(self, input_shape, num_outputs):super(Dueling_DQN, self).__init__()self.input_shape = input_shapeself.num_actions = num_outputsself.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)

改進究竟管用與否？

下面是DDQN（藍色），Dueling DQN（粉紅）和DQN（橘黃）在Pong上面得表現(xiàn)。可以看到DDQN > Dueling DQN > DQN，其中DDQN大約比DQN得收斂快10%。

總結(jié)

以上是生活随笔為你收集整理的从DQN到Double DQN和Dueling DQN——pytorch实操的全部內(nèi)容，希望文章能夠幫你解決所遇到的問題。

如果覺得生活随笔網(wǎng)站內(nèi)容還不錯，歡迎將生活随笔推薦給好友。