DQN的学习效果还是很惊艳的,首先放上本次实验的代码。和官方给出的例子一样,是托举平衡杆的问题。
给出视频链接:强化学习DQN
import torch
import torch.nn as nn
import torch.nn.functional as F
import numpy as np
from torch.autograd import Variable
import gym
# 超参数
BATCH_SIZE = 32
LR = 0.01 # learning rate
# 强化学习的参数
EPSILON = 0.9 # greedy policy
GAMMA = 0.9 # reward discount
TARGET_REPLACE_ITER = 200 # target update frequency
MEMORY_CAPACITY = 2000
# 导入实验环境
env = gym.make('CartPole-v0')
env = env.unwrapped
N_ACTIONS = env.action_space.n
N_STATES = env.observation_space.shape[0]
class Net(nn.Module):
def __init__(self, ):
super(Net, self).__init__()
self.fc1 = nn.Linear(N_STATES, 10)
self.fc1.weight.data.normal_(0, 0.1) # 初始化
self.out = nn.Linear(10, N_ACTIONS)
self.out.weight.data.normal_(0, 0.1) # 初始化
def forward(self, x):
x = self.fc1(x)
x = F.relu(x)
actions_value = self.out(x)
return actions_value
class DQN(object):
def __init__(self):
self.eval_net, self.target_net = Net(), Net()
# 记录学习到多少步
self.learn_step_counter = 0 # for target update
self.memory_counter = 0 # for storing memory
# 初始化memory
self.memory = np.zeros((MEMORY_CAPACITY, N_STATES * 2 + 2))
self.optimizer = torch.optim.Adam(self.eval_net.parameters(), lr=LR)
self.loss_func = nn.MSELoss()
def choose_action(self, x):
x = Variable(torch.unsqueeze(torch.FloatTensor(x), 0))
if np.random.uniform() < EPSILON:
action_value = self.eval_net.forward(x)
action = torch.max(action_value, 1)[1].data.numpy()[0]
else: # random
action = np.random.randint(0, N_ACTIONS)
return action
# s:当前状态, a:动作, r:reward奖励, s_:下一步状态
def store_transaction(self, s, a, r, s_):
transaction = np.hstack((s, [a, r], s_))
# replace the old memory with new memory
index = self.memory_counter % MEMORY_CAPACITY
self.memory[index, :] = transaction
self.memory_counter += 1
def learn(self):
# target net update
if self.learn_step_counter % TARGET_REPLACE_ITER == 0:
self.target_net.load_state_dict(self.eval_net.state_dict())
self.learn_step_counter += 1
sample_index = np.random.choice(MEMORY_CAPACITY, BATCH_SIZE) #从2000中随机抽取32个样本
b_memory = self.memory[sample_index, :] #通过索引找到对应的值
b_s = Variable(torch.FloatTensor(b_memory[:, :N_STATES]))
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: ]))
q_eval = self.eval_net(b_s).gather(1, b_a)
q_next = self.target_net(b_s_).detach()
# here = q_next.max(1)[0]
# there = torch.reshape(here,(32,1))
# print(there)
# there = torch.transpose(here,0,1)
q_target = b_r + GAMMA * q_next.max(1)[0].reshape((32,1))
# here = here*GAMMA
loss = self.loss_func(q_eval, q_target)
self.optimizer.zero_grad()
loss.backward()
self.optimizer.step()
dqn = DQN()
print('\nCollecting experience...')
for i_episode in range(4000):
s = env.reset()
while True:
env.render()
a = dqn.choose_action(s)
# take action
s_, r, done, info = env.step(a)
# modify the reward
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
dqn.store_transaction(s, a, r, s_)
print(str(dqn.memory_counter) +'\t'+ str(i_episode))
if dqn.memory_counter > MEMORY_CAPACITY:
dqn.learn()
if done:
break
s = s_
- DQN如上文所说,采用的是两个神经网络,在程序中为
eval_net
和target_net
,两个神经网络具有相同的结构,不同之处是一个实时更新权重,一个间隔一定周期更新权重TARGET_REPLACE_ITER
。 - DQN的精髓之处在于 Experience learning 和 Fix-Q target。首先,程序刚开始运行的时候,前2000步都是在积累经验,这个地方和记忆池是一个容量大小。也就是说前2000步,eval_net是一直都有输出的,给定state,输出action。然后通过Q-learning的思想进行贪婪选择,执行。而 target_net此时没有参与(有权重初值的)。当2000步骤之后,开始学习,也就是开始更新神经网络了。
- 更新神经网络的时候,先从 记忆池 中随机抽取一定的记忆(batch = 32),然后用s输入给 eval_net,将 s_ 输入给target_net。得到 q_eval 和 q_next, 然后计算loss, 反向更新神经网络权重。这里需要注意的是,target_net 并不参与权重的更新,仍然保留着之前的参数,等到一定的间隔时候,才进行神经网络权值的直接copy。
- 每次训练神经网络的时候都取出一批数据进行训练,一方面随机取出,大大减少了数据之间的相关性,另一方面,加快了训练的步伐,效果还是挺明显的。