虚拟主机网站500错误高端网站建设报价

一、代码

# 4-2 用ε贪婪算法解决多老虎机问题

import numpy as np
import matplotlib.pyplot as plt

# 参数设置
c = 10  # 老虎机数量
l = 501  # 最后时刻
runs = 1000  # 运行次数
epsilon = 0.01  # ε贪婪中的ε

# 构造随机种子为指定值的随机数生成器
rng = np.random.default_rng(1)

# 初始化每一时刻的奖赏和
timestep_rewards = np.zeros(l)  # 修改为大小为 l 的数组
timestep_rewards_epsilon = np.zeros(l)  # 修改为大小为 l 的数组

# 运行循环
for run in range(runs):
    # 每次运行的初始化
    occ_actions = np.zeros(c)  # 每个行动被选择的次数
    acc_rewards = np.zeros(c)  # 每个行动下累加获得的奖赏
    estimated_rewards = np.zeros(c)  # 每台老虎机奖赏期望值的估计值
    occ_actions_epsilon = np.zeros(c)  # 每个行动被选择的次数（ε贪婪）
    acc_rewards_epsilon = np.zeros(c)  # 每个行动下累加获得的奖赏（ε贪婪）
    estimated_rewards_epsilon = np.zeros(c)  # 每台老虎机奖赏期望值的估计值（ε贪婪）
    means_bandits = rng.normal(0, 1, c)  # 每次运行时的老虎机奖赏期望值

    # 从第1个时刻到第l-1个时刻
    for t in range(l - 1):
        # 计算每台老虎机的奖赏期望值的估计值
        for i in range(c):
            estimated_rewards[i] = 0 if acc_rewards[i] == 0 else acc_rewards[i] / occ_actions[i]
            estimated_rewards_epsilon[i] = 0 if acc_rewards_epsilon[i] == 0 else acc_rewards_epsilon[i] / occ_actions_epsilon[i]

        # 选择贪婪行动
        a_t = np.argmax(estimated_rewards).item()
        # ε贪婪
        if rng.random() > epsilon:
            # 选择贪婪行动
            a_t_epsilon = np.argmax(estimated_rewards_epsilon).item()
        else:
            # 随机选择行动
            a_t_epsilon = rng.integers(0, c)

        # 选择该行动后（在下一时刻）获得的奖赏
        r_tp1 = rng.normal(means_bandits[a_t], 1)
        r_tp1_epsilon = rng.normal(means_bandits[a_t_epsilon], 1)

        # 累加当前行动选择获得的奖赏
        occ_actions[a_t] += 1
        acc_rewards[a_t] += r_tp1
        occ_actions_epsilon[a_t_epsilon] += 1
        acc_rewards_epsilon[a_t_epsilon] += r_tp1_epsilon

        # 累加每个时刻下获得的奖赏
        timestep_rewards[t + 1] += r_tp1
        timestep_rewards_epsilon[t + 1] += r_tp1_epsilon

# 画平均奖赏线
plt.figure()
plt.plot(np.arange(1, l), timestep_rewards[1:] / runs, linewidth=2)
plt.plot(np.arange(1, l), timestep_rewards_epsilon[1:] / runs, 'r', linewidth=2)
plt.ylabel('Averaged rewards')
plt.xlabel('Time steps')
plt.title('Greedy v.s. Epsilon-greedy')
plt.legend(['Greedy', 'Epsilon-greedy'])
plt.show()

二、结果图