简介

机器学习：监督学习、非监督学习、强化学习

• 模仿人类和动物的试错机制进行学习
• 智能体与环境交互，根据当前的环境状态s，按照一定策略采取行动a，获得回报r
• 目标：获取最大累积期望回报

脉络介绍：

gym库-CliffWalking

安装标准化实验环境

pip install gym

CliffWalking：悬崖寻路问题，4*12网格，红色为悬崖，36为起始，47为目标

动作：0-4，上右下左，如果移出除网络则不变
奖励：{-1,100}，与悬崖为-100，否则为-1

SARSA

行动选择策略：ε-greedy，以ε的概率进行探索，以1-ε的概率进行利用
Q值更新：根据下一次实际行动更新，胆小，选择离悬崖远的路线

 td_target += gamma * Q[next_state, next_action]
 Q[state, action] += lr * (td_target - Q[state, action])

Q-learning

行动选择策略：ε-greedy，以ε的概率进行探索，以1-ε的概率进行利用
Q值更新：选取最优的行动更新Q值，胆大，最终选择离悬崖近的路线

 td_target += gamma * max(Q[next_state, :])
 Q[state, action] += lr * (td_target - Q[state, action])

示例

SARSA

import numpy as np
import pandas as pd
import gym
from tqdm import tqdm


def max_index(a):
    # return np.argmax(a)
    candidate = np.where(a == np.max(a))[0]  
    index = np.random.randint(0, len(candidate))  
    return candidate[index]


def eps_greedy(Q, state):
    a = Q[state, :]
    if np.random.random() < 1 - eps:
        return max_index(a)
    return np.random.randint(0, len(a))  # [start,end)


def calc_policy(Q):
    state_number = Q.shape[0]
    policy = np.zeros(shape=state_number, dtype=np.int8)
    for i in range(state_number):
        policy[i] = np.argmax(Q[i, :])
    return policy


# 0123：{上右下左}
def print_optimal_action(pi, row, col):
    print(actions)
    for i in range(row):
        for j in range(col):
            print(actions[pi[i * col + j]], end=' ')
        print()


# 比较a,b之间的差值是否小于阈值
def is_same(a, b, thresold=0.001):
    e = np.abs(a - b) > thresold
    return np.sum(e) == 0


if __name__ == '__main__':
    eps = 0.01
    lr = 0.01
    gamma = 0.99

    np.set_printoptions(suppress=True)
    row, col = 4, 12
    state_number = row * col
    action_number = 4
    actions = list('↑→↓←')  # 上右下左：0123
    Q = np.zeros((state_number, action_number), dtype=np.float64)
    Q_last = [np.ones_like(Q), np.ones_like(Q), np.ones_like(Q), np.ones_like(Q), np.ones_like(Q)]
    env = gym.make('CliffWalking-v0')
    print('状态数量：', env.observation_space.n)
    print('行为数量：', env.action_space.n)

    #  10万个episode的迭代
    for i in tqdm(range(1, 100000)):
        env.reset()
        state = 36  # 初始位置
        done = False
        action = eps_greedy(Q, state)
        while not done:
            next_state, reward, done, info = env.step(action)
            # print('state, action, reward:', state, action, reward)
            next_action = eps_greedy(Q, next_state)
            td_target = reward
            if not done:
                td_target += gamma * Q[next_state, next_action]
            Q[state, action] += lr * (td_target - Q[state, action])
            state = next_state
            action = next_action
        if is_same(Q_last[0], Q):
            print('Q-table迭代完成，提前退出：', i)
            break
        Q_last = Q_last[1:]
        Q_last.append(np.copy(Q))
    pi = calc_policy(Q)
    print('Q Table:\n', Q)
    # np.savetxt('Q_table.txt', Q, fmt='%.5f')
    pd.DataFrame(Q).to_excel('Q_table.xlsx', index=True)
    for s_id in range(state_number):
        print(s_id, s_id // col, s_id % col, Q[s_id, :], pi[s_id], actions[pi[s_id]])
    print('最优策略：', pi)
    print_optimal_action(pi, row, col)

    # 输出最终路径（状态及坐标）
    env.reset()
    state = 36
    done = False
    trace = [{state: (state // col, state % col)}]
    while not done:
        action = pi[state]
        state, _, done, _ = env.step(action)
        trace.append({state: (state // col, state % col)})
        if len(trace) > 48:
            break
    print(len(trace), '：', trace)

Q-learning

import numpy as np
import pandas as pd
import gym
from tqdm import tqdm


def max_index(a):
    candidate = np.where(a == np.max(a))[0]
    index = np.random.randint(0, len(candidate))
    return candidate[index]


def eps_greedy(Q, state):
    a = Q[state, :]
    if np.random.random() < 1-eps:
        return max_index(a)
    return np.random.randint(0, len(a))


def calc_policy(Q):
    state_number = Q.shape[0]
    policy = np.zeros(shape=state_number, dtype=np.int8)
    for i in range(state_number):
        policy[i] = np.argmax(Q[i, :])
    return policy


# 0123：{上右下左}
def print_optimal_action(pi, row, col):
    a = list('↑→↓←')
    print(a)
    for i in range(row):
        for j in range(col):
            print(a[pi[i*col+j]], end=' ')
        print()


if __name__ == '__main__':
    eps = 0.1 # 10%概率探索，90%概率利用
    lr = 0.01
    gamma = 0.99

    np.set_printoptions(suppress=True)

    row = 4
    col = 12
    state_number = row * col
    action_number = 4   # 上下左右
    Q = np.zeros((state_number, action_number), dtype=np.float64)
    env = gym.make('CliffWalking-v0')
    for i in tqdm(range(10000)):   # 10000个episode的训练
        env.reset()
        state = 36  
        done = False
        action = eps_greedy(Q, state)
        while not done:
            next_state, reward, done, info = env.step(action)
            next_action = eps_greedy(Q, next_state)
            td_target = reward
            if not done:
                td_target += gamma * max(Q[next_state, :])  # Q-learning
            Q[state, action] += lr * (td_target - Q[state, action])
            state = next_state
            action = next_action
    pi = calc_policy(Q)
    print('Q Table:\n', Q)
    pd.DataFrame(Q).to_excel('Q_table.xlsx', index=True)
    for s_id in range(state_number):
        print(s_id, s_id//col, s_id % col, Q[s_id, :], pi[s_id])
    print_optimal_action(pi, row, col)

    # 输出最终路径
    env.reset()
    state = 36
    done = False
    trace = [{state: (state // col, state % col)}]
    while not done:
        action = np.argmax(Q[state, :])
        state, _, done, _ = env.step(action)
        trace.append({state: (state//col, state % col)})
        if len(trace) > 48:
            break
    print(len(trace), '：', trace)