整体代码如下：
python
import gym
import numpy as np
import torch
import matplotlib.pyplot as plt
from tqdm import tqdm
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
print(device)
def moving_average(a, window_size):
cumulative_sum = np.cumsum(np.insert(a, 0, 0))
middle = (cumulative_sum[window_size:] - cumulative_sum[:-window_size]) / window_size
r = np.arange(1, window_size-1, 2)
begin = np.cumsum(a[:window_size-1])[::2] / r
end = (np.cumsum(a[:-window_size:-1])[::2] / r)[::-1]
return np.concatenate((begin, middle, end))
class PolicyNetwork(torch.nn.Module):
def __init__(self,statedim,hiddendim,actiondim):
super(PolicyNetwork,self).__init__()
self.cf1=torch.nn.Linear(statedim,hiddendim)
self.cf2=torch.nn.Linear(hiddendim,actiondim)
def forward(self,x):
x=torch.nn.functional.relu(self.cf1(x))
return torch.nn.functional.softmax(self.cf2(x),dim=1)
class REINFORCE:
def __init__(self,statedim,hiddendim,actiondim,learningrate,gamma,device):
self.policynet=PolicyNetwork(statedim,hiddendim,actiondim).to(device)
self.gamma=gamma
self.device=device
self.optimizer=torch.optim.Adam(self.policynet.parameters(),lr=learningrate)
def takeaction(self,state):
state=torch.tensor([state],dtype=torch.float).to(self.device)
probs=self.policynet(state)
actiondist=torch.distributions.Categorical(probs)#torch.distributions.Categorical：这是 PyTorch 中用于表示类别分布的类,可以使用 actiondist.sample() 方法从这个分布中随机采样一个类别
action=actiondist.sample()
return action.item()
def update(self,transitiondist):
statelist=transitiondist['states']
rewardlist=transitiondist['rewards']
actionlist=transitiondist['actions']
G=0
self.optimizer.zero_grad()
for i in reversed(range(len(rewardlist))):#从最后一步计算起
reward=rewardlist[i]
state=statelist[i]
action=actionlist[i]
state=torch.tensor([state],dtype=torch.float).to(self.device)
action=torch.tensor([action]).view(-1,1).to(self.device)
logprob=torch.log(self.policynet(state).gather(1,action)) #.gather(1, action) 方法从策略网络的输出中提取对应于特定动作 action 的概率值。这里的 1 表示沿着维度 1（通常对应于动作维度）进行索引。
G=self.gamma*G+reward
loss=-logprob*G#每一步的损失函数
loss.backward()#反向传播计算梯度
self.optimizer.step()#更新参数,梯度下降
learningrate=4e-3
episodesnum=1000
hiddendim=128
gamma=0.99
pbarnum=10
printreturnnum=10
device=torch.device('cuda') if torch.cuda.is_available() else torch.device('cpu')
env=gym.make('CartPole-v1')
env.reset(seed=880)
torch.manual_seed(880)
statedim=env.observation_space.shape[0]
actiondim=env.action_space.n
agent=REINFORCE(statedim=statedim,hiddendim=hiddendim,actiondim=actiondim,learningrate=learningrate,gamma=gamma,device=device)
returnlist=[]
for k in range(pbarnum):
with tqdm(total=int(episodesnum/pbarnum),desc='Iteration %d'%k)as pbar:
for episode in range(int(episodesnum/pbarnum)):
g=0
transitiondist={'states':[],'actions':[],'nextstates':[],'rewards':[]}
state,_=env.reset(seed=880)
done=False
while not done:
action=agent.takeaction(state)
nextstate,reward,done,truncated,_=env.step(action)
done=done or truncated
transitiondist['states'].append(state)
transitiondist['actions'].append(action)
transitiondist['nextstates'].append(nextstate)
transitiondist['rewards'].append(reward)
state=nextstate
g=g+reward
returnlist.append(g)
agent.update(transitiondist)
if (episode+1)%(printreturnnum)==0:
pbar.set_postfix({'Episode':'%d'%(episodesnum//pbarnum+episode+1),'Return':'%.3f'%np.mean(returnlist[-printreturnnum:])})
pbar.update(1)
episodelist=list(range(len(returnlist)))
plt.plot(episodelist,returnlist)
plt.xlabel('Episodes')
plt.ylabel('Returns')
plt.title('REINFORCE on {}'.format(env.spec.name))
plt.show()
mvreturn=moving_average(returnlist,9)
plt.plot(episodelist,mvreturn)
plt.xlabel('Episodes')
plt.ylabel('Returns')
plt.title('REINFORCE on {}'.format(env.spec.name))
plt.show()
效果：
