注意
单击此处下载完整的示例代码
训练玩马里奥的 RL 代理¶
创建时间: 2020年12月17日 |上次更新时间:2024 年 2 月 5 日 |上次验证时间:未验证
作者:Yuansong Feng, Suraj Subramanian, Howard Wang, Steven Guo.
本教程将引导您了解 Deep Reinforcement 的基础知识 学习。最后,您将实现一个 AI 驱动的 Mario(使用 Double Deep Q-Networks),它 可以单独玩游戏。
尽管本教程不需要 RL 的先验知识,但您需要 可以熟悉这些 RL 概念, 并拥有这份方便的作弊单作为您的伴侣。此处提供了完整的代码。
%%bash
pip install gym-super-mario-bros==7.4.0
pip install tensordict==0.3.0
pip install torchrl==0.3.0
import torch
from torch import nn
from torchvision import transforms as T
from PIL import Image
import numpy as np
from pathlib import Path
from collections import deque
import random, datetime, os
# Gym is an OpenAI toolkit for RL
import gym
from gym.spaces import Box
from gym.wrappers import FrameStack
# NES Emulator for OpenAI Gym
from nes_py.wrappers import JoypadSpace
# Super Mario environment for OpenAI Gym
import gym_super_mario_bros
from tensordict import TensorDict
from torchrl.data import TensorDictReplayBuffer, LazyMemmapStorage
RL 定义¶
环境代理与之交互并从中学习的世界。
操作 \(a\):代理如何响应环境。这 所有可能的 Actions 的集合称为 action-space。
状态 \(s\) :环境的当前特征。这 Environment 可以处于的所有可能状态的集合称为 State-space。
奖励 \(r\) : 奖励是从环境到 代理。这是驱使 Agent 学习和改变其未来的原因 行动。多个时间步长的奖励聚合称为 Return。
最优动作值函数 \(Q^*(s,a)\) : 给出期望的 返回 如果你从状态 \(s\) 开始,则执行任意操作 \(a\),然后对于每个未来的时间步长,执行 实现回报最大化。\(Q\) 可以说代表 state 中的操作。我们尝试近似此函数。
环境¶
初始化环境¶
在马里奥中,环境由管子、蘑菇和其他 组件。
当 Mario 执行操作时,环境会使用更改的 (next) 状态、奖励和其他信息。
# Initialize Super Mario environment (in v0.26 change render mode to 'human' to see results on the screen)
if gym.__version__ < '0.26':
env = gym_super_mario_bros.make("SuperMarioBros-1-1-v0", new_step_api=True)
else:
env = gym_super_mario_bros.make("SuperMarioBros-1-1-v0", render_mode='rgb', apply_api_compatibility=True)
# Limit the action-space to
# 0. walk right
# 1. jump right
env = JoypadSpace(env, [["right"], ["right", "A"]])
env.reset()
next_state, reward, done, trunc, info = env.step(action=0)
print(f"{next_state.shape},\n {reward},\n {done},\n {info}")
/usr/local/lib/python3.10/dist-packages/gym/envs/registration.py:555: UserWarning:
WARN: The environment SuperMarioBros-1-1-v0 is out of date. You should consider upgrading to version `v3`.
/usr/local/lib/python3.10/dist-packages/gym/envs/registration.py:627: UserWarning:
WARN: The environment creator metadata doesn't include `render_modes`, contains: ['render.modes', 'video.frames_per_second']
/usr/local/lib/python3.10/dist-packages/gym/utils/passive_env_checker.py:233: DeprecationWarning:
`np.bool8` is a deprecated alias for `np.bool_`. (Deprecated NumPy 1.24)
(240, 256, 3),
0.0,
False,
{'coins': 0, 'flag_get': False, 'life': 2, 'score': 0, 'stage': 1, 'status': 'small', 'time': 400, 'world': 1, 'x_pos': 40, 'y_pos': 79}
预处理环境¶
环境数据将返回到 中的代理。如你所见
在上面,每个状态都由一个 size 数组表示。
通常,这比我们的代理需要的信息更多;例如
马里奥的行动不取决于管道的颜色或天空!next_state
[3, 240, 256]
我们使用 Wrapper 对环境数据进行预处理,然后再将其发送到 代理。
GrayScaleObservation
是转换 RGB 图像的常用包装器
到灰度;这样做会减小 state 表示的大小
而不会丢失有用的信息。现在每个状态的大小:[1, 240, 256]
ResizeObservation
将每个观测值缩减为方形图像。
新尺寸:[1, 84, 84]
SkipFrame
是一个自定义包装器,它继承自 和
实现该函数。因为连续帧不会
变化很大,我们可以跳过 n 个中间帧而不会损失太多
信息。第 n 帧汇总了每帧累积的奖励
跳过的帧。gym.Wrapper
step()
FrameStack
是一个包装器,允许我们压缩连续的帧
将环境转化为单个观测点,以馈送到我们的
学习模型。这样,我们可以确定马里奥是着陆还是
根据他前几次的移动方向进行跳跃
框架。
class SkipFrame(gym.Wrapper):
def __init__(self, env, skip):
"""Return only every `skip`-th frame"""
super().__init__(env)
self._skip = skip
def step(self, action):
"""Repeat action, and sum reward"""
total_reward = 0.0
for i in range(self._skip):
# Accumulate reward and repeat the same action
obs, reward, done, trunk, info = self.env.step(action)
total_reward += reward
if done:
break
return obs, total_reward, done, trunk, info
class GrayScaleObservation(gym.ObservationWrapper):
def __init__(self, env):
super().__init__(env)
obs_shape = self.observation_space.shape[:2]
self.observation_space = Box(low=0, high=255, shape=obs_shape, dtype=np.uint8)
def permute_orientation(self, observation):
# permute [H, W, C] array to [C, H, W] tensor
observation = np.transpose(observation, (2, 0, 1))
observation = torch.tensor(observation.copy(), dtype=torch.float)
return observation
def observation(self, observation):
observation = self.permute_orientation(observation)
transform = T.Grayscale()
observation = transform(observation)
return observation
class ResizeObservation(gym.ObservationWrapper):
def __init__(self, env, shape):
super().__init__(env)
if isinstance(shape, int):
self.shape = (shape, shape)
else:
self.shape = tuple(shape)
obs_shape = self.shape + self.observation_space.shape[2:]
self.observation_space = Box(low=0, high=255, shape=obs_shape, dtype=np.uint8)
def observation(self, observation):
transforms = T.Compose(
[T.Resize(self.shape, antialias=True), T.Normalize(0, 255)]
)
observation = transforms(observation).squeeze(0)
return observation
# Apply Wrappers to environment
env = SkipFrame(env, skip=4)
env = GrayScaleObservation(env)
env = ResizeObservation(env, shape=84)
if gym.__version__ < '0.26':
env = FrameStack(env, num_stack=4, new_step_api=True)
else:
env = FrameStack(env, num_stack=4)
将上述包装器应用于环境后,最终包装的
state 由 4 个灰度缩放的连续帧堆叠在一起组成,因为
如上图左图所示。每次 Mario 执行操作时,
环境使用此结构的状态进行响应。结构
由大小为 的三维数组表示。[4, 84, 84]
代理¶
我们创建一个类来表示游戏中的代理。马里奥
应该能够:Mario
根据基于当前 状态(环境)。
记住经验。经验 = (当前状态,当前 action、reward、next state)。马里奥缓存并随后召回他的 体验更新其操作策略。
随着时间的推移,了解更好的操作策略
class Mario:
def __init__():
pass
def act(self, state):
"""Given a state, choose an epsilon-greedy action"""
pass
def cache(self, experience):
"""Add the experience to memory"""
pass
def recall(self):
"""Sample experiences from memory"""
pass
def learn(self):
"""Update online action value (Q) function with a batch of experiences"""
pass
在以下部分中,我们将填充 Mario 的参数和 定义他的职能。
做¶
对于任何给定状态,代理都可以选择执行最佳操作 (利用)或随机操作 (探索)。
马里奥随机探索,有几率 ;什么时候
他选择利用,他依赖 (在本节中实现) 来提供最佳操作。self.exploration_rate
MarioNet
Learn
class Mario:
def __init__(self, state_dim, action_dim, save_dir):
self.state_dim = state_dim
self.action_dim = action_dim
self.save_dir = save_dir
self.device = "cuda" if torch.cuda.is_available() else "cpu"
# Mario's DNN to predict the most optimal action - we implement this in the Learn section
self.net = MarioNet(self.state_dim, self.action_dim).float()
self.net = self.net.to(device=self.device)
self.exploration_rate = 1
self.exploration_rate_decay = 0.99999975
self.exploration_rate_min = 0.1
self.curr_step = 0
self.save_every = 5e5 # no. of experiences between saving Mario Net
def act(self, state):
"""
Given a state, choose an epsilon-greedy action and update value of step.
Inputs:
state(``LazyFrame``): A single observation of the current state, dimension is (state_dim)
Outputs:
``action_idx`` (``int``): An integer representing which action Mario will perform
"""
# EXPLORE
if np.random.rand() < self.exploration_rate:
action_idx = np.random.randint(self.action_dim)
# EXPLOIT
else:
state = state[0].__array__() if isinstance(state, tuple) else state.__array__()
state = torch.tensor(state, device=self.device).unsqueeze(0)
action_values = self.net(state, model="online")
action_idx = torch.argmax(action_values, axis=1).item()
# decrease exploration_rate
self.exploration_rate *= self.exploration_rate_decay
self.exploration_rate = max(self.exploration_rate_min, self.exploration_rate)
# increment step
self.curr_step += 1
return action_idx
缓存和调用¶
这两个函数充当 Mario 的“记忆”过程。
cache()
:每次马里奥执行一个动作时,他都会存储到他的记忆中。他的经验包括当前状态、执行的操作、操作的奖励、下一个状态、
以及游戏是否完成。experience
recall()
:Mario 从他的
memory,并使用它来学习游戏。
class Mario(Mario): # subclassing for continuity
def __init__(self, state_dim, action_dim, save_dir):
super().__init__(state_dim, action_dim, save_dir)
self.memory = TensorDictReplayBuffer(storage=LazyMemmapStorage(100000, device=torch.device("cpu")))
self.batch_size = 32
def cache(self, state, next_state, action, reward, done):
"""
Store the experience to self.memory (replay buffer)
Inputs:
state (``LazyFrame``),
next_state (``LazyFrame``),
action (``int``),
reward (``float``),
done(``bool``))
"""
def first_if_tuple(x):
return x[0] if isinstance(x, tuple) else x
state = first_if_tuple(state).__array__()
next_state = first_if_tuple(next_state).__array__()
state = torch.tensor(state)
next_state = torch.tensor(next_state)
action = torch.tensor([action])
reward = torch.tensor([reward])
done = torch.tensor([done])
# self.memory.append((state, next_state, action, reward, done,))
self.memory.add(TensorDict({"state": state, "next_state": next_state, "action": action, "reward": reward, "done": done}, batch_size=[]))
def recall(self):
"""
Retrieve a batch of experiences from memory
"""
batch = self.memory.sample(self.batch_size).to(self.device)
state, next_state, action, reward, done = (batch.get(key) for key in ("state", "next_state", "action", "reward", "done"))
return state, next_state, action.squeeze(), reward.squeeze(), done.squeeze()
学习¶
Mario 在后台使用 DDQN 算法。DDQN 使用两个卷积网络 - \(Q_{online}\) 和 \(Q_{target}\) - 它们独立地近似于最优 action-value 函数。
在我们的实现中,我们在 \(Q_{online}\) 和 \(Q_{target}\) 之间共享特征生成器,但保持单独的 FC
分类器。\(\theta_{target}\) (\(Q_{target}\) 的参数)被冻结以防止反向传播更新。相反
它会定期与 \(\theta_{online}\) 同步(更多相关信息
稍后)。features
神经网络¶
class MarioNet(nn.Module):
"""mini CNN structure
input -> (conv2d + relu) x 3 -> flatten -> (dense + relu) x 2 -> output
"""
def __init__(self, input_dim, output_dim):
super().__init__()
c, h, w = input_dim
if h != 84:
raise ValueError(f"Expecting input height: 84, got: {h}")
if w != 84:
raise ValueError(f"Expecting input width: 84, got: {w}")
self.online = self.__build_cnn(c, output_dim)
self.target = self.__build_cnn(c, output_dim)
self.target.load_state_dict(self.online.state_dict())
# Q_target parameters are frozen.
for p in self.target.parameters():
p.requires_grad = False
def forward(self, input, model):
if model == "online":
return self.online(input)
elif model == "target":
return self.target(input)
def __build_cnn(self, c, output_dim):
return nn.Sequential(
nn.Conv2d(in_channels=c, out_channels=32, kernel_size=8, stride=4),
nn.ReLU(),
nn.Conv2d(in_channels=32, out_channels=64, kernel_size=4, stride=2),
nn.ReLU(),
nn.Conv2d(in_channels=64, out_channels=64, kernel_size=3, stride=1),
nn.ReLU(),
nn.Flatten(),
nn.Linear(3136, 512),
nn.ReLU(),
nn.Linear(512, output_dim),
)
TD估计和TD目标¶
学习涉及两个值:
TD Estimate - 给定状态 \(s\) 的预测最差 \(Q^*\)
TD Target - 当前奖励和下一个状态 \(s'\) 的估计 \(Q^*\) 的聚合
因为我们不知道下一步 \(a'\) 会是什么,所以我们使用 动作 \(A'\) 使下一个状态 \(s'\) 中的 \(Q_{online}\) 最大化。
请注意,我们使用 @torch.no_grad() 装饰器来禁用此处的梯度计算
(因为我们不需要在 \(\theta_{target}\) 上反向传播)。td_target()
class Mario(Mario):
def __init__(self, state_dim, action_dim, save_dir):
super().__init__(state_dim, action_dim, save_dir)
self.gamma = 0.9
def td_estimate(self, state, action):
current_Q = self.net(state, model="online")[
np.arange(0, self.batch_size), action
] # Q_online(s,a)
return current_Q
@torch.no_grad()
def td_target(self, reward, next_state, done):
next_state_Q = self.net(next_state, model="online")
best_action = torch.argmax(next_state_Q, axis=1)
next_Q = self.net(next_state, model="target")[
np.arange(0, self.batch_size), best_action
]
return (reward + (1 - done.float()) * self.gamma * next_Q).float()
更新模型¶
当马里奥从他的重放缓冲区中采样输入时,我们计算 \(TD_t\) 和 \(TD_e\),并将这个损失向下反向传播 \(Q_{online}\) 到
更新其参数 \(\theta_{online}\) (\(\alpha\) 是
传递给lr
optimizer
)
\(\theta_{target}\) 不会通过反向传播进行更新。 相反,我们会定期将 \(\theta_{online}\) 复制到 \(\theta_{target}\)
class Mario(Mario):
def __init__(self, state_dim, action_dim, save_dir):
super().__init__(state_dim, action_dim, save_dir)
self.optimizer = torch.optim.Adam(self.net.parameters(), lr=0.00025)
self.loss_fn = torch.nn.SmoothL1Loss()
def update_Q_online(self, td_estimate, td_target):
loss = self.loss_fn(td_estimate, td_target)
self.optimizer.zero_grad()
loss.backward()
self.optimizer.step()
return loss.item()
def sync_Q_target(self):
self.net.target.load_state_dict(self.net.online.state_dict())
保存检查点¶
class Mario(Mario):
def save(self):
save_path = (
self.save_dir / f"mario_net_{int(self.curr_step // self.save_every)}.chkpt"
)
torch.save(
dict(model=self.net.state_dict(), exploration_rate=self.exploration_rate),
save_path,
)
print(f"MarioNet saved to {save_path} at step {self.curr_step}")
把它们放在一起¶
class Mario(Mario):
def __init__(self, state_dim, action_dim, save_dir):
super().__init__(state_dim, action_dim, save_dir)
self.burnin = 1e4 # min. experiences before training
self.learn_every = 3 # no. of experiences between updates to Q_online
self.sync_every = 1e4 # no. of experiences between Q_target & Q_online sync
def learn(self):
if self.curr_step % self.sync_every == 0:
self.sync_Q_target()
if self.curr_step % self.save_every == 0:
self.save()
if self.curr_step < self.burnin:
return None, None
if self.curr_step % self.learn_every != 0:
return None, None
# Sample from memory
state, next_state, action, reward, done = self.recall()
# Get TD Estimate
td_est = self.td_estimate(state, action)
# Get TD Target
td_tgt = self.td_target(reward, next_state, done)
# Backpropagate loss through Q_online
loss = self.update_Q_online(td_est, td_tgt)
return (td_est.mean().item(), loss)
伐木¶
import numpy as np
import time, datetime
import matplotlib.pyplot as plt
class MetricLogger:
def __init__(self, save_dir):
self.save_log = save_dir / "log"
with open(self.save_log, "w") as f:
f.write(
f"{'Episode':>8}{'Step':>8}{'Epsilon':>10}{'MeanReward':>15}"
f"{'MeanLength':>15}{'MeanLoss':>15}{'MeanQValue':>15}"
f"{'TimeDelta':>15}{'Time':>20}\n"
)
self.ep_rewards_plot = save_dir / "reward_plot.jpg"
self.ep_lengths_plot = save_dir / "length_plot.jpg"
self.ep_avg_losses_plot = save_dir / "loss_plot.jpg"
self.ep_avg_qs_plot = save_dir / "q_plot.jpg"
# History metrics
self.ep_rewards = []
self.ep_lengths = []
self.ep_avg_losses = []
self.ep_avg_qs = []
# Moving averages, added for every call to record()
self.moving_avg_ep_rewards = []
self.moving_avg_ep_lengths = []
self.moving_avg_ep_avg_losses = []
self.moving_avg_ep_avg_qs = []
# Current episode metric
self.init_episode()
# Timing
self.record_time = time.time()
def log_step(self, reward, loss, q):
self.curr_ep_reward += reward
self.curr_ep_length += 1
if loss:
self.curr_ep_loss += loss
self.curr_ep_q += q
self.curr_ep_loss_length += 1
def log_episode(self):
"Mark end of episode"
self.ep_rewards.append(self.curr_ep_reward)
self.ep_lengths.append(self.curr_ep_length)
if self.curr_ep_loss_length == 0:
ep_avg_loss = 0
ep_avg_q = 0
else:
ep_avg_loss = np.round(self.curr_ep_loss / self.curr_ep_loss_length, 5)
ep_avg_q = np.round(self.curr_ep_q / self.curr_ep_loss_length, 5)
self.ep_avg_losses.append(ep_avg_loss)
self.ep_avg_qs.append(ep_avg_q)
self.init_episode()
def init_episode(self):
self.curr_ep_reward = 0.0
self.curr_ep_length = 0
self.curr_ep_loss = 0.0
self.curr_ep_q = 0.0
self.curr_ep_loss_length = 0
def record(self, episode, epsilon, step):
mean_ep_reward = np.round(np.mean(self.ep_rewards[-100:]), 3)
mean_ep_length = np.round(np.mean(self.ep_lengths[-100:]), 3)
mean_ep_loss = np.round(np.mean(self.ep_avg_losses[-100:]), 3)
mean_ep_q = np.round(np.mean(self.ep_avg_qs[-100:]), 3)
self.moving_avg_ep_rewards.append(mean_ep_reward)
self.moving_avg_ep_lengths.append(mean_ep_length)
self.moving_avg_ep_avg_losses.append(mean_ep_loss)
self.moving_avg_ep_avg_qs.append(mean_ep_q)
last_record_time = self.record_time
self.record_time = time.time()
time_since_last_record = np.round(self.record_time - last_record_time, 3)
print(
f"Episode {episode} - "
f"Step {step} - "
f"Epsilon {epsilon} - "
f"Mean Reward {mean_ep_reward} - "
f"Mean Length {mean_ep_length} - "
f"Mean Loss {mean_ep_loss} - "
f"Mean Q Value {mean_ep_q} - "
f"Time Delta {time_since_last_record} - "
f"Time {datetime.datetime.now().strftime('%Y-%m-%dT%H:%M:%S')}"
)
with open(self.save_log, "a") as f:
f.write(
f"{episode:8d}{step:8d}{epsilon:10.3f}"
f"{mean_ep_reward:15.3f}{mean_ep_length:15.3f}{mean_ep_loss:15.3f}{mean_ep_q:15.3f}"
f"{time_since_last_record:15.3f}"
f"{datetime.datetime.now().strftime('%Y-%m-%dT%H:%M:%S'):>20}\n"
)
for metric in ["ep_lengths", "ep_avg_losses", "ep_avg_qs", "ep_rewards"]:
plt.clf()
plt.plot(getattr(self, f"moving_avg_{metric}"), label=f"moving_avg_{metric}")
plt.legend()
plt.savefig(getattr(self, f"{metric}_plot"))
让我们玩吧!¶
在此示例中,我们运行了 40 集的训练循环,但为了让 Mario 真正学习 他的世界,我们建议至少运行 40,000 集的循环!
use_cuda = torch.cuda.is_available()
print(f"Using CUDA: {use_cuda}")
print()
save_dir = Path("checkpoints") / datetime.datetime.now().strftime("%Y-%m-%dT%H-%M-%S")
save_dir.mkdir(parents=True)
mario = Mario(state_dim=(4, 84, 84), action_dim=env.action_space.n, save_dir=save_dir)
logger = MetricLogger(save_dir)
episodes = 40
for e in range(episodes):
state = env.reset()
# Play the game!
while True:
# Run agent on the state
action = mario.act(state)
# Agent performs action
next_state, reward, done, trunc, info = env.step(action)
# Remember
mario.cache(state, next_state, action, reward, done)
# Learn
q, loss = mario.learn()
# Logging
logger.log_step(reward, loss, q)
# Update state
state = next_state
# Check if end of game
if done or info["flag_get"]:
break
logger.log_episode()
if (e % 20 == 0) or (e == episodes - 1):
logger.record(episode=e, epsilon=mario.exploration_rate, step=mario.curr_step)
Using CUDA: True
Episode 0 - Step 163 - Epsilon 0.9999592508251706 - Mean Reward 635.0 - Mean Length 163.0 - Mean Loss 0.0 - Mean Q Value 0.0 - Time Delta 1.958 - Time 2025-01-02T22:15:50
Episode 20 - Step 5007 - Epsilon 0.9987490329557962 - Mean Reward 667.429 - Mean Length 238.429 - Mean Loss 0.0 - Mean Q Value 0.0 - Time Delta 56.842 - Time 2025-01-02T22:16:47
Episode 39 - Step 8854 - Epsilon 0.9977889477081997 - Mean Reward 656.6 - Mean Length 221.35 - Mean Loss 0.0 - Mean Q Value 0.0 - Time Delta 45.301 - Time 2025-01-02T22:17:32
结论¶
在本教程中,我们了解了如何使用 PyTorch 训练玩游戏的 AI。您可以使用相同的方法 训练 AI 在 OpenAI 健身房玩任何游戏。希望您喜欢本教程,请随时通过我们的 github 与我们联系!
脚本总运行时间:(1 分 45.054 秒)