今天看别的的代码中用到了 numpy.random.seed()，网上查阅后说是使用后可以在下一次生成相同的随机数，但这到底有什么用呢。
我的理解是两种程序对比，都是随机开始，这时要设置seed，让两种程序的初始条件相同，更加有利于对比，下面上代码：

  import numpy as np
import matplotlib.pyplot as plt

arms = 10
# 设置seed
seed = np.random.randint(0, 100)

# max = 2.24
variance = 1
mean = np.zeros(10)
Q_table = np.zeros(10)
Q_steps = np.zeros(10)

total_steps = 20000
avg_reward = 0.0
step = 0

def walk_mean(mean):
    for i in xrange(len(mean)):
        if np.random.random() > 0.5:
            mean[i] += 1
        else:
            mean[i] -= 1
    return mean

def greedy_update(avg_reward, step, Q_table, mean):
    mean = walk_mean(mean)
    choice = np.argmax(Q_table)
    step += 1
    reward = np.random.normal(mean[choice], variance)
    avg_reward += (reward-avg_reward)/step
    Q_steps[choice] += 1
    Q_table[choice] += (reward-Q_table[choice])/Q_steps[choice]
    return avg_reward, step, Q_table, mean

def fix_greedy_update(avg_reward, step, Q_table, mean):
    mean = walk_mean(mean)
    choice = np.argmax(Q_table)
    step += 1
    reward = np.random.normal(mean[choice], variance)
    avg_reward += (reward-avg_reward)/step
    Q_steps[choice] += 1
    Q_table[choice] += (reward-Q_table[choice])/10
    return avg_reward, step, Q_table, mean

def e_greedy_update(avg_reward, step, Q_table, p, mean):
    flag = np.random.random()
    if flag <= p:
        mean = walk_mean(mean)
        choice = np.random.randint(0, 10)
        step += 1
        reward = np.random.normal(mean[choice], variance)
        avg_reward += (reward-avg_reward)/step
        Q_steps[choice] += 1
        Q_table[choice] += (reward-Q_table[choice])/Q_steps[choice]
    else:
        avg_reward, step, Q_table, mean = greedy_update(avg_reward, step, Q_table, mean)
    return avg_reward, step, Q_table, mean

def fix_e_greedy_update(avg_reward, step, Q_table, p, mean):
    flag = np.random.random()
    if flag <= p:
        mean = walk_mean(mean)
        choice = np.random.randint(0, 10)
        step += 1
        reward = np.random.normal(mean[choice], variance)
        avg_reward += (reward-avg_reward)/step
        Q_table[choice] += (reward-Q_table[choice])/10
    else:
        avg_reward, step, Q_table, mean = fix_greedy_update(avg_reward, step, Q_table, mean)
    return avg_reward, step, Q_table, mean

# e-greedy with e=0.1
np.random.seed(seed)
r = []
for i in xrange(total_steps):
    avg_reward, step, Q_table, mean = e_greedy_update(avg_reward, step, Q_table, 0.1, mean)
    r.append(avg_reward)
print Q_table
print Q_steps
one, = plt.plot(r, label='e_greedy_0.1')

# fix-e-greedy with e=0.1
Q_table = np.zeros(10)
Q_steps = np.zeros(10)
mean = np.zeros(10)
avg_reward = 0.0
step = 0
np.random.seed(seed)
r = []
for i in xrange(total_steps):
    avg_reward, step, Q_table, mean = fix_e_greedy_update(avg_reward, step, Q_table, 0.1, mean)
    r.append(avg_reward)
print Q_table
print Q_steps
two, = plt.plot(r, label='fix_e_greedy_0.1')


plt.legend(handles=[one, two])

plt.show()

可以看到，在 # e-greedy with e=0.1 与 # fix-e-greedy with e=0.1下面分别调用了 np.random.seed(seed) ,这样使得 两种对比程序的初始条件相同：flag = np.random.random()。