基于强化学习的gym Mountain Car稳定控制

基于强化学习的gym Mountain Car稳定控制

依赖包版本

gym == 0.21.0
stable-baselines3 == 1.6.2

环境测试

环境介绍：Mountain Car

import gym


# Create environment
env = gym.make("MountainCar-v0")

eposides = 10
for eq in range(eposides):
    obs = env.reset()
    done = False
    rewards = 0
    while not done:
        action = env.action_space.sample()
        obs, reward, done, info = env.step(action)
        env.render()
        rewards += reward
    print(rewards)

环境测试视频：Mountain Car test

Q-learning 模型

模型训练

import gym
import numpy as np

env = gym.make("MountainCar-v0")

# Q-Learning settings
LEARNING_RATE = 0.1
DISCOUNT = 0.95
EPISODES = 25000

SHOW_EVERY = 1000

# Exploration settings
epsilon = 1  # not a constant, qoing to be decayed
START_EPSILON_DECAYING = 1
END_EPSILON_DECAYING = EPISODES//2
epsilon_decay_value = epsilon/(END_EPSILON_DECAYING - START_EPSILON_DECAYING)

DISCRETE_OS_SIZE = [20, 20]
discrete_os_win_size = (env.observation_space.high - env.observation_space.low) / DISCRETE_OS_SIZE

print(discrete_os_win_size)


def get_discrete_state(state):
    discrete_state = (state - env.observation_space.low)/discrete_os_win_size
    return tuple(discrete_state.astype(np.int64))  # we use this tuple to look up the 3 Q values for the available actions in the q-


q_table = np.random.uniform(low=-2, high=0, size=(DISCRETE_OS_SIZE + [env.action_space.n]))


for episode in range(EPISODES):
    state = env.reset()
    discrete_state = get_discrete_state(state)

    if episode % SHOW_EVERY == 0:
        render = True
        print(episode)
    else:
        render = False

    done = False
    while not done:
        if np.random.random() > epsilon:
            # Get action from Q table
            action = np.argmax(q_table[discrete_state])
        else:
            # Get random action
            action = np.random.randint(0, env.action_space.n)

        new_state, reward, done, _ = env.step(action)
        new_discrete_state = get_discrete_state(new_state)

        # If simulation did not end yet after last step - update Q table
        if not done:
            # Maximum possible Q value in next step (for new state)
            max_future_q = np.max(q_table[new_discrete_state])
            # Current Q value (for current state and performed action)
            current_q = q_table[discrete_state + (action,)]
            # And here's our equation for a new Q value for current state and action
            new_q = (1 - LEARNING_RATE) * current_q + LEARNING_RATE * (reward + DISCOUNT * max_future_q)
            # Update Q table with new Q value
            q_table[discrete_state + (action,)] = new_q
            # Simulation ended (for any reson) - if goal position is achived - update Q value with reward directly

        elif new_state[0] >= env.goal_position:
            # q_table[discrete_state + (action,)] = reward
            q_table[discrete_state + (action,)] = 0
            print("we made it on episode {}".format(episode))

        discrete_state = new_discrete_state

        if render:
            env.render()

    # Decaying is being done every episode if episode number is within decaying range
    if END_EPSILON_DECAYING >= episode >= START_EPSILON_DECAYING:
        epsilon -= epsilon_decay_value

np.save("q_table.npy", arr=q_table)

env.close()

鉴于该环境相对简单，用Q-learning方法生成的q表较小，不存在维度爆炸问题，采用Q-learning能实现准确控制，读取生成的q表测试代码如下：

import gym
import numpy as np


env = gym.make("MountainCar-v0")

# Q-Learning settings
LEARNING_RATE = 0.1
DISCOUNT = 0.95
EPISODES = 10

DISCRETE_OS_SIZE = [20, 20]
discrete_os_win_size = (env.observation_space.high - env.observation_space.low) / DISCRETE_OS_SIZE

def get_discrete_state(state):
    discrete_state = (state - env.observation_space.low)/discrete_os_win_size
    return tuple(discrete_state.astype(np.int64))  # we use this tuple to look up the 3 Q values for the available actions in the q-

q_table = np.load(file="q_table.npy")

for episode in range(EPISODES):
    state = env.reset()
    discrete_state = get_discrete_state(state)

    rewards = 0
    done = False
    while not done:
        # Get action from Q table
        action = np.argmax(q_table[discrete_state])
        new_state, reward, done, _ = env.step(action)
        new_discrete_state = get_discrete_state(new_state)

        rewards += reward

        # If simulation did not end yet after last step - update Q table
        if done and new_state[0] >= env.goal_position:
            print("we made it on episode {}, rewards {}".format(episode, rewards))

        discrete_state = new_discrete_state
        env.render()

env.close()

Q-learning测试视频结果：Mountain Car Qlearning
由视频可以看出，小车每次都能够达到目标点。

DQN模型

采用stable-baseline3默认的DQN网络架构（64，64），学习率为5e-4，训练次数为1.5million次，训练代码如下：

import gym
from stable_baselines3 import DQN


# Create environment
env = gym.make("MountainCar-v0")

model = DQN(
    "MlpPolicy",
    env,
    verbose=1,
    learning_rate=5e-4)

# Train the agent and display a progress bar
model.learn(
    total_timesteps=int(1.5e6),
    progress_bar=True)

# Save the agent
model.save("DQN_MountainCar")

模型测试代码如下：

import gym
from stable_baselines3 import DQN
from stable_baselines3.common.evaluation import evaluate_policy


# Create environment
env = gym.make("MountainCar-v0")

# load model
model = DQN.load("DQN_MountainCar", env=env)

mean_reward, std_reward = evaluate_policy(
    model,
    model.get_env(),
    deterministic=True,
    render=True,
    n_eval_episodes=10)
print(mean_reward)

测试结果视频：Mountain Car DQN
根据视频可看出小车每次都能到达终点。

后记

stable-baseline3: 手册
gym: 手册