运行效果

我手动最多打出10分，模型能打30多……

问题分析

• 时间不连续，最小单位为“帧”
• 状态status是连续的浮点数值
• 动作action只有2种，即“升”和“不升”，无论采取什么动作，都作用于玩家加速度而非直接改变位置
• 除了刚刚通过门时采取的动作外，动作的结果全是必然没有随机性

与玩家决策相关的量很多比如门（腔缝）的高度和宽度，飞机本身尺寸等等，具体要什么不要什么：

• 玩家左侧与门右侧水平距离占当前两门间距的比例
• 玩家中点与门中点垂直距离占总高度的比例
• 玩家y向速度与“最大速度”的比例
• 玩家中点与屏幕水平中线的距离占屏幕高度的一半的比例

我们最后用这3个量作为模型入参，所以模型输入3通道但输出只有2通道。只要玩家存活就得到正奖励。

环境搭建

• 安装CUDA
• CUDA版本：nvcc --version

  nvcc: NVIDIA (R) Cuda compiler driver
  Copyright (c) 2005-2021 NVIDIA Corporation
  Built on Sun_Feb_14_21:12:58_PST_2021
  Cuda compilation tools, release 11.2, V11.2.152
  Build cuda_11.2.r11.2/compiler.29618528_0
  

• Python版本：python --version

  Python 3.9.12
  

• 系统版本：cat /proc/version

  Linux version 5.4.0-109-generic (buildd@ubuntu) (gcc version 9.4.0 (Ubuntu 9.4.0-1ubuntu1~20.04.1)) #123-Ubuntu SMP Fri Apr 8 09:10:54 UTC 2022
  

• 安装依赖

  pip install pygame autopep8 numpy
  pip install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu113
  

目录结构

• 📁 assets
  • 📁 textures
    • 🖼 door.gif
    • 🖼 player_age0.gif
    • 🖼 player_age1.gif
• 📄 main.py
• 📄 game.py
• 📄 util.py

代码

main.py

# -*- coding: utf-8 -*-
"""训练和预测。
"""
import random
import sys
from collections import OrderedDict

import numpy as np
import pygame
import torch
from torch import nn, optim

from game import Game
from util import print_bar


class Model(nn.Module):
    """Dueling DQN结构。
    """

    def __init__(self):
        super(Model, self).__init__()
        self.layers = nn.ModuleDict({
            'c': nn.Sequential(nn.Linear(3, 12, device=CUDA), nn.Sigmoid()),
            'a': nn.Linear(12, 2, device=CUDA),
            'v': nn.Linear(12, 1, device=CUDA),
            'o': nn.ReLU(),
        })

    def forward(self, arg: torch.Tensor) -> torch.Tensor:
        """模型前向传播。

        Parameters
        ----------
        x : torch.Tensor
            样本输入模型

        Returns
        -------
        torch.Tensor
            预测值。
        """
        output = arg
        output = self.layers['c'](output)
        adv = self.layers['a'](output)
        val = self.layers['v'](output)
        output = self.layers['o'](adv+val)
        return output

    def load_params(self, model: 'Model', rate: float = 1):
        """模型参数软更新。

        Parameters
        ----------
        model : Model
            将这个模型的参数复制到当前模型
        rate : float, optional
            `1`表示将模型参数完全复制到当前模型, by default 1
        """
        for key, value in self.layers.items():
            if rate >= 1.:
                forign = model.layers[key].state_dict()
                value.load_state_dict(forign)
            else:
                local = value.state_dict()
                forign = model.layers[key].state_dict()
                mix = OrderedDict()
                for key in local.keys():
                    mix[key] = local.get(key)*(1-rate) + forign.get(key)*rate
                value.load_state_dict(mix)


def simulate(model: Model, batch_size: int, epslion: float = .1, eval_step: int = None, env_args: dict = None) -> 'tuple[list,float,int]':
    """模拟游戏过程并收集数据。

    Parameters
    ----------
    model : Model
        决策用
    batch_size : int
        收集数据总条数
    epslion : float, optional
        尝试比例, by default .1
    eval_step : int, optional
        模型将控制游戏的最大步数，参与模型评估, by default `batch_size`
    env_args : dict, optional
        环境初始化参数, by default None

    Returns
    -------
    tuple[list,float,int]
        采集的数据, 平均存活时长, 无探索情况下生存时间
    """
    cache = []
    env = Game(**env_args, without_screen=True)
    livetimes = []
    livetime = 0
    for _ in range(batch_size):
        state = env.shot()
        if random.random() <= epslion:
            action_index = random.randint(0, len(ACTIONS)-1)
        else:
            values = model(torch.tensor(state, device=CUDA))
            action_index = values.argmax(-1)
        jump = ACTIONS[action_index]
        env.step(jump)
        next_state = env.shot()
        reward = float(env.playing)
        cache.append((state, action_index, next_state, reward))
        if not env.playing:
            env = Game(**env_args, without_screen=True)
            livetimes.append(livetime)
        else:
            livetime += 1
    env = Game(**env_args, without_screen=True)
    max_step = eval_step or batch_size
    livetime = 0
    for _ in range(max_step):  # 看模型在不进行随机探索条件下能维持多少帧不摔机，这是评估标准
        state = env.shot()
        values = model(torch.tensor(state, device=CUDA))
        action_index = values.argmax(-1)
        jump = ACTIONS[action_index]
        env.step(jump)
        if not env.playing:
            break
        livetime += 1
    return cache, sum(livetimes)/max(1, len(livetimes))/batch_size, livetime


def train(policy_net: Model, opt: optim.Optimizer, loss_func: 'nn._Loss', epochs: int, batch_size: int, cache_size: int, epslion: float = .1, gamma: float = .5, update_ratio: float = .5, eval_step: int = None, target_accuracy=.99, env_args: dict = None) -> 'tuple[Model,list[float],list[float],list[int]]':
    """训练模型。

    Parameters
    ----------
    policy_net : Model
        决策网络对象
    opt : optim.Optimizer
        优化器
    loss_func : nn._Loss
        损失函数
    epochs : int
        迭代轮数
    batch_size : int
        批量
    epslion : float, optional
        探索动作比例, by default .1
    gamma : float, optional
        未来奖励权重，`0`表示仅考虑当前奖励, by default .5
    update_ratio : float, optional
        软更新比例, by default .5
    target_accuracy : float, optional
        模型决策目标得分, by default .99
    env_args : dict, optional
        环境初始化参数, by default None

    Returns
    -------
    tuple[Model,list[float],list[float],list[int]]
        目标网络, 损失, 存活时间
    """
    target_net = Model()
    target_net.load_params(policy_net)
    policy_net.train(mode=True)
    target_net.train(mode=False)
    loss_vals, accuracies, livetimes, cache = [], [], [], []
    for epoch in range(epochs):
        target_net.load_params(policy_net, update_ratio)

        # 获取数据
        batch, accuracy, livetime = simulate(model=target_net, batch_size=batch_size, epslion=epslion, eval_step=eval_step, env_args=env_args)
        accuracies.append(accuracy)
        livetimes.append(livetime)
        if livetime/(eval_step or batch_size) >= target_accuracy:
            # 模型的决策已经达标不需要再训练了
            break
        # 装入经验池
        cache.extend(batch)
        cache = cache[-cache_size:]

        # 经验池抽样并转换成tensor
        states, actions, nexts, rewards = [], [], [], []
        for state, action, next_state, reward in random.sample(cache, batch_size):
            states.append(state)
            actions.append(action)
            rewards.append(reward)
            nexts.append(next_state)
        states = torch.tensor(states, device=CUDA)
        actions = torch.tensor(actions, device=CUDA).unsqueeze(-1)
        rewards = torch.tensor(rewards, device=CUDA)
        nexts = torch.tensor(nexts, device=CUDA)

        # 计算输出与损失，批量梯度下降
        v_target = target_net.forward(nexts).detach()
        y_target = v_target.max(dim=-1).values * gamma
        y_target += rewards * (1-gamma)
        v_eval = policy_net.forward(states)
        y_eval = v_eval.gather(index=actions, dim=-1)
        loss = loss_func(y_eval, y_target)
        opt.zero_grad()
        loss.backward()
        opt.step()

        loss = loss.item()
        loss_vals.append(loss)
        print_bar(epoch+1, epochs, ("%.10f" % loss, '%.10f' % accuracy, livetime))
    return target_net, loss_vals, accuracies, livetimes


np.set_printoptions(suppress=True)
CUDA = torch.device("cuda")
MODEL = Model()
OPT = optim.Adam(MODEL.parameters(), lr=.01)
LOSS_FUNCTION = nn.MSELoss()
ACTIONS = (True, False)
SCREEN_SIZE = (800, 600)
FPS = 20
GAME_CONFIG = {
    'screen_size': SCREEN_SIZE,
    'door_size': (80, 180),
    'speed': 10,
    'jump_force': 3,
    'g': 2,
    'door_distance': 60,
}
if __name__ == "__main__":
    pygame.init()  # 初始化
    model, loss_vals, accuracies, livetimes = train(
        policy_net=MODEL,
        opt=OPT,
        loss_func=LOSS_FUNCTION,
        epochs=20000,
        batch_size=192,
        cache_size=2000,
        epslion=.3,
        gamma=.9,
        update_ratio=.1,
        target_accuracy=.95,
        env_args=GAME_CONFIG,
        eval_step=1200,
    )

    # 使用模型决策并观看结果
    print('\n\n')
    model = model.to('cpu')
    model.train(mode=False)
    SCREEN = pygame.display.set_mode(SCREEN_SIZE)
    fcclock = pygame.time.Clock()
    game = Game(**GAME_CONFIG)
    while True:
        # 循环，直到接收到窗口关闭事件
        for event in pygame.event.get():
            # 处理事件
            if event.type == pygame.QUIT:
                # 接收到窗口关闭事件
                pygame.quit()
                sys.exit()
        keys = pygame.key.get_pressed()
        if keys[pygame.K_ESCAPE]:
            pygame.quit()
            sys.exit()
        else:
            state = torch.tensor(game.shot())
            values = model.forward(state)
            action_index = values.argmax(-1)
            jump = ACTIONS[action_index]
        game.step(jump)
        pygame.display.set_caption(f'SCORE: {game.score}')
        game.draw(SCREEN)
        fcclock.tick(FPS)
        pygame.display.update()
        if not game.playing:
            # 自动开局
            game = Game(**GAME_CONFIG)

game.py

# -*- coding: utf-8 -*-
"""游戏环境相关。
"""
import random
import sys
import pygame


class Box:
    """包含基础位置、尺寸、速度、加速度的盒子类。
    """
    __position = None
    __size = None
    __speed = None
    __acceleration = None

    def __init__(self, cx: int, cy: int, w: int, h: int, sx: int = 0, sy: int = 0, ax: int = 0, ay: int = 0):
        self.__position = [cx, cy]
        self.__size = [w, h]
        self.__speed = [sx or 0, sy or 0]
        self.__acceleration = [ax or 0, ay or 0]

    @property
    def width(self):
        return self.__size[0]

    @property
    def height(self):
        return self.__size[-1]

    @property
    def size(self):
        return self.__size

    @property
    def x(self):
        return self.__position[0]

    @property
    def y(self):
        return self.__position[-1]

    @property
    def position(self):
        return self.__position

    @property
    def speed_x(self):
        return self.__speed[0]

    @speed_x.setter
    def speed_x(self, v):
        self.__speed[0] = v

    @property
    def speed_y(self):
        return self.__speed[-1]

    @speed_y.setter
    def speed_y(self, v):
        self.__speed[-1] = v

    @property
    def speed(self):
        return self.__speed

    @speed.setter
    def speed(self, v: 'tuple[int,int]'):
        self.__speed[0] = v[0]
        self.__speed[-1] = v[-1]

    @property
    def acceleration_x(self):
        return self.__acceleration[0]

    @acceleration_x.setter
    def acceleration_x(self, v: int):
        self.__acceleration[0] = v

    @property
    def acceleration_y(self):
        return self.__acceleration[-1]

    @acceleration_y.setter
    def acceleration_y(self, v: int):
        self.__acceleration[-1] = v

    @property
    def acceleration(self):
        return self.__acceleration

    @acceleration.setter
    def acceleration(self, v: 'tuple[int,int]'):
        self.__acceleration[0] = v[0]
        self.__acceleration[-1] = v[-1]

    @property
    def left(self):
        return self.x-self.width/2

    @property
    def right(self):
        return self.x+self.width/2

    @property
    def top(self):
        return self.y-self.height/2

    @property
    def bottom(self):
        return self.y+self.height/2

    def move(self, force_x: int = None, force_y: int = None):
        """为盒子施力使其移动。

        Parameters
        ----------
        force_x : int, optional
            水平分量, by default None
        force_y : int, optional
            垂直分量, by default None
        """
        self.acceleration_x = force_x or 0
        self.acceleration_y = force_y or 0
        self.speed_x += self.acceleration_x
        self.speed_y += self.acceleration_y
        self.__position[0] += self.speed_x
        self.__position[-1] += self.speed_y


def is_intersect(player: Box, door: Box) -> bool:
    return (door.top > player.top or player.bottom > door.bottom) \
        and not (player.left >= door.right or door.left >= player.right)


class GameObject(Box):
    """游戏基础对象。
    """

    def __init__(self, imgs: list, img_cd: int = 1, *args, **kwargs):
        super(GameObject, self).__init__(*args, **kwargs)
        self.__imgs = [item for item in imgs]
        self.__img_cd = img_cd or -1
        self.living = True
        self.img_index = -1

    def img_grow(self):
        self.img_index = (self.img_index+1) % self.__img_cd

    @property
    def img(self):
        return self.__imgs[self.img_index]


class Game:
    door_size = None
    player = None
    jump_force = 0
    g = 1
    door_distance = 0
    doors = None
    time = 1
    score = 0

    def __init__(self, screen_size=(800, 600), player_size=(160, 80), door_size=(80, 160), speed=5, jump_force=1.3, g=0.4, door_distance=100, max_falling_speed: int = 100, without_screen=False, **_):
        self.player = GameObject(
            cx=screen_size[0]/4,
            cy=screen_size[1]/2,
            w=player_size[0],
            h=player_size[1],
            sx=0, sy=0,
            ax=0, ay=g,
            imgs=[None, ] if without_screen else[
                pygame.image.load('./assets/textures/player_age0.gif').convert_alpha(),
                pygame.image.load('./assets/textures/player_age1.gif').convert_alpha(),
            ],
            img_cd=2
        )
        self.without_screen = without_screen
        self.screen_size = screen_size
        self.door_size = door_size
        self.speed = speed
        self.jump_force = jump_force
        self.g = g
        self.door_distance = door_distance
        self.max_falling_speed = max_falling_speed
        self.doors = [self.create_door()]

    @property
    def playing(self) -> bool:
        """描述玩家是否存活。
        """
        return self.player.living

    @property
    def door(self) -> 'GameObject|None':
        """距离玩家最近的且玩家未穿过的门。
        """
        for door in self.doors:
            if door.right >= self.player.left:
                return door
        return None

    def create_door(self) -> GameObject:
        """随机初始化门。

        Returns
        -------
        GameObject
            屏幕右侧随机位置的门。
        """
        door = GameObject(
            cx=self.screen_size[0]+self.door_size[0]/2,
            cy=random.randint(self.door_size[1]/2, self.screen_size[1]-self.door_size[1]/2),
            w=self.door_size[0],
            h=self.door_size[1],
            sx=-self.speed,
            imgs=[None, ] if self.without_screen else [pygame.image.load('./assets/textures/door.gif').convert_alpha(),],
            img_cd=2
        )
        return door

    def draw(self, surface: 'pygame.Surface'):
        """绘制游戏帧。

        Parameters
        ----------
        surface : pygame.Surface
            pygame屏幕
        """
        if not self.player.living:
            return
        surface.fill([86, 92, 66])
        self.player.img_grow()
        surface.blit(pygame.transform.scale(self.player.img, (self.player.width, self.player.height)), (self.player.left, self.player.top))
        for door in self.doors:
            surface.blit(pygame.transform.scale(door.img, (door.width, door.top)), (door.left, 0))
            surface.blit(pygame.transform.scale(door.img, (door.width, self.screen_size[1]-door.bottom)), (door.left, door.bottom))

    @staticmethod
    def __shot(door: Box, player: Box, screen_size: 'tuple[int,int]', speed_scale: int) -> 'list[float]':
        return [(door.right-player.left)/screen_size[0], (player.y-door.y)/screen_size[-1], player.speed_y/speed_scale, ]

    def shot(self) -> 'list[float]':
        """组装并返回当前游戏环境状态。

        Returns
        -------
        list[float]
            模型所需的多元组。
        """
        return Game.__shot( self.door, self.player, [self.door_distance*self.speed, self.screen_size[-1]], self.max_falling_speed, )

    def step(self, jump: 'bool|int|float' = False):
        """游戏步进。

        Parameters
        ----------
        jump : bool, optional
            玩家是否跳跃, by default False
        """
        # 玩家必须存活才能继续游戏
        if not self.player.living:
            return

        if self.time % self.door_distance == 0 or not (self.doors and len(self.doors)):
            # 时间间隔生成门，时间重置
            self.doors.append(self.create_door())
            self.time = 1
        else:
            # 时间正常递增直到时间间隔
            self.time += 1

        # 清除已经移除屏幕的门
        while self.doors[0].right < 0:
            del self.doors[0]

        # 移动玩家和所有门
        for door in self.doors:
            door.move()
        door = self.door
        living = 0 < self.player.y < self.screen_size[1] and not is_intersect(self.player, door)
        self.player.move(None, -self.jump_force if jump else self.g)
        if jump:
            self.player.speed_y = min(0, self.player.speed_y)

        self.player.living = living
        # 判断玩家和门存活
        if door.living and self.player.left >= door.right:
            door.living = False
            self.score += 1

util.py

# -*- coding: utf-8 -*-
"""输出打印工具模块。
"""

def print_bar(epoch, epochs, etc=None, bar_size=50):
    """打印进度条。

    Parameters
    ----------
    epoch : int
        当前进度
    epochs : int
        总进度
    etc : Any, optional
        打印后缀, by default None
    bar_size : int, optional
        进度条长度, by default 50
    """
    process = bar_size*epoch/epochs
    process = int(process+(int(process) < process))
    strs = [
        f"Epoch {epoch}/{epochs}",
        f" |\033[1;30;47m{' ' * process}\033[0m{' ' * (bar_size-process)}| ",
    ]
    if etc is not None:
        strs.append(str(etc))
    if epoch:
        strs.insert(0, "\033[A")
    print("".join(strs)+"    ")

door.gif

player_age0.gif

player_age1.gif