策略梯度玩 cartpole 游戏，强化学习代替PID算法控制平衡杆

cartpole游戏，车上顶着一个自由摆动的杆子，实现杆子的平衡，杆子每次倒向一端车就开始移动让杆子保持动态直立的状态，策略函数使用一个两层的简单神经网络，输入状态有4个，车位置，车速度，杆角度，杆速度，输出action为左移动或右移动，输入状态发现至少要给3个才能稳定一会儿，给2个完全学不明白，给4个能学到很稳定的policy

策略梯度实现代码，使用torch实现一个简单的神经网络

import gym
import torch
import torch.nn as nn
import torch.optim as optim
import pygame
import sys
from collections import deque
import numpy as np

# 策略网络定义
class PolicyNetwork(nn.Module):
    def __init__(self):
        super(PolicyNetwork, self).__init__()
        self.fc = nn.Sequential(
            nn.Linear(4, 10),  # 4个状态输入，128个隐藏单元
            nn.Tanh(),
            nn.Linear(10, 2),  # 输出2个动作的概率
            nn.Softmax(dim=-1)
        )

    def forward(self, x):
        # print(x)  车位置 车速度 杆角度 杆速度
        selected_values = x[:, [0,1,2,3]]  #只使用车位置和杆角度
        return self.fc(selected_values)

# 训练函数
def train(policy_net, optimizer, trajectories):
    policy_net.zero_grad()
    loss = 0
    print(trajectories[0])
    for trajectory in trajectories:
        
        # if trajectory["returns"] > 90:
        # returns = torch.tensor(trajectory["returns"]).float()
        # else:
        returns = torch.tensor(trajectory["returns"]).float() - torch.tensor(trajectory["step_mean_reward"]).float()
        # print(f"获得奖励{returns}")
        log_probs = trajectory["log_prob"]
        loss += -(log_probs * returns).sum()  # 计算策略梯度损失
    loss.backward()
    optimizer.step()
    return loss.item()

# 主函数
def main():
    env = gym.make('CartPole-v1')
    policy_net = PolicyNetwork()
    optimizer = optim.Adam(policy_net.parameters(), lr=0.01)

    print(env.action_space)
    print(env.observation_space)
    pygame.init()
    screen = pygame.display.set_mode((600, 400))
    clock = pygame.time.Clock()

    rewards_one_episode= []
    for episode in range(10000):
        
        state = env.reset()
        done = False
        trajectories = []
        state = state[0]
        step = 0
        torch.save(policy_net, 'policy_net_full.pth')
        while not done:
            state_tensor = torch.tensor(state).float().unsqueeze(0)
            probs = policy_net(state_tensor)
            action = torch.distributions.Categorical(probs).sample().item()
            log_prob = torch.log(probs.squeeze(0)[action])
            next_state, reward, done, _,_ = env.step(action)

            # print(episode)
            trajectories.append({"state": state, "action": action, "reward": reward, "log_prob": log_prob})
            state = next_state

            for event in pygame.event.get():
                if event.type == pygame.QUIT:
                    pygame.quit()
                    sys.exit()
            step +=1
            
            # 绘制环境状态
            if rewards_one_episode and rewards_one_episode[-1] >99:
                screen.fill((255, 255, 255))
                cart_x = int(state[0] * 100 + 300)
                pygame.draw.rect(screen, (0, 0, 255), (cart_x, 300, 50, 30))
                # print(state)
                pygame.draw.line(screen, (255, 0, 0), (cart_x + 25, 300), (cart_x + 25 - int(50 * torch.sin(torch.tensor(state[2]))), 300 - int(50 * torch.cos(torch.tensor(state[2])))), 2)
                pygame.display.flip()
                clock.tick(200)
                

        print(f"第{episode}回合",f"运行{step}步后挂了")
        # 为策略梯度计算累积回报
        returns = 0
        
        
        for traj in reversed(trajectories):
            returns = traj["reward"] + 0.99 * returns
            traj["returns"] = returns
            if rewards_one_episode:
                # print(rewards_one_episode[:10])
                traj["step_mean_reward"] = np.mean(rewards_one_episode[-10:])
            else:
                traj["step_mean_reward"] = 0
        rewards_one_episode.append(returns)
        # print(rewards_one_episode[:10])
        train(policy_net, optimizer, trajectories)

def play():

    env = gym.make('CartPole-v1')
    policy_net = PolicyNetwork()
    pygame.init()
    screen = pygame.display.set_mode((600, 400))
    clock = pygame.time.Clock()

    state = env.reset()
    done = False
    trajectories = deque()
    state = state[0]
    step = 0
    policy_net = torch.load('policy_net_full.pth')
    while not done:
        state_tensor = torch.tensor(state).float().unsqueeze(0)
        probs = policy_net(state_tensor)
        action = torch.distributions.Categorical(probs).sample().item()
        log_prob = torch.log(probs.squeeze(0)[action])
        next_state, reward, done, _,_ = env.step(action)

        # print(episode)
        trajectories.append({"state": state, "action": action, "reward": reward, "log_prob": log_prob})
        state = next_state

        for event in pygame.event.get():
            if event.type == pygame.QUIT:
                pygame.quit()
                sys.exit()

        
        # 绘制环境状态
        screen.fill((255, 255, 255))
        cart_x = int(state[0] * 100 + 300)
        pygame.draw.rect(screen, (0, 0, 255), (cart_x, 300, 50, 30))
        # print(state)
        pygame.draw.line(screen, (255, 0, 0), (cart_x + 25, 300), (cart_x + 25 - int(50 * torch.sin(torch.tensor(state[2]))), 300 - int(50 * torch.cos(torch.tensor(state[2])))), 2)
        pygame.display.flip()
        clock.tick(60)
        step +=1

    print(f"运行{step}步后挂了")



if __name__ == '__main__':
    main() #训练
    # play() #推理

运行效果，训练过程不是很稳定，有时候学很多轮次也学不明白，有时侯只需要几十次就可以学明白了