python用softmax算法实现多摇臂老虎机强化学习应该怎么写？

Question

我在网上找到了用贪心算法实现多摇臂老虎机强化学习的代码，把贪心算法换成了softmax算法，但是关于老虎机各个摇臂被选中的概率有一些疑问，代码如下：

import numpy as np
import random

#老虎机的定义
class Bandit:

    def __init__(self, arms_prob):
        self.arms_prob = arms_prob
        self.size = len(self.arms_prob)

    def play(self, i):
        if not 0 <= i < self.size:
            return -1
        else:
            if random.uniform(0, 1) < self.arms_prob[i]:
                return 1
            else:
                return 0

class Model:
    def __init__(self, bandit, temperature=0.2, training_epochs=10000):
        self.bandit = bandit #摇臂老虎机
        self.temperature = temperature #温度参数
        self.training_epochs = training_epochs #尝试次数
        self.values = np.zeros(bandit.size) #摇臂的平均奖赏
        self.times = np.zeros(bandit.size) #尝试次数
        self.p = [1/bandit.size for i in range(bandit.size)] #摇臂的概率分布
        self.size = bandit.size #摇臂数
        self.result = 0 #回报率最高的摇臂编号

    #计算摇臂选择的概率分布
    def probability(self,index):
        self.a = 0
        for i in range(bandit.size):
            self.a += np.e ** (self.values[i] / self.temperature)
        return (np.e ** (self.values[index] / self.temperature)) / self.a

    def train(self):
        i = 0
        r = 0 #最终的奖赏
        while i <= self.training_epochs:
            index = random.choices([i for i in range(bandit.size)],self.p)[0]#根据权重选取摇臂
            reward = self.bandit.play(index)
            assert reward >= 0 and index >= 0
            self.times[index] += 1
            self.values[index] = 1.0 / (self.times[index]+1) *(self.values[index] * self.times[index] + reward)
            for j in range(bandit.size):   #就是这三行
                self.p[j] = self.probability(j)#就是这三行
            # self.p[index] = self.probability(index)#就是这三行
            i += 1
            print("Round %d, choose %d, reward %d " % (i, index, reward))
            print(repr(self.values))
            r += reward #累加奖赏
            # print(repr(self.p))
            sum_ = 0
            for z in range (self.size):
                sum_ += self.p[z]
            print(sum_)

        self.result = self.values.argmax(axis=0)
        print('累积奖赏：',r,' 回报率：', r / self.training_epochs)

bandit = Bandit([0.5, 0.6, 0.8, 0.9, 0.3, 0.95, 0.96, 0.45, 0.93, 0.22, 0.65])
model = Model(bandit, temperature=0.1, training_epochs=30000)
# bandit = Bandit([0.4, 0.2])
# model = Model(bandit, temperature=0.01, training_epochs=3000)
model.train()
print("最佳选择是 %d." % model.result)

在47-49的三行里，我写了两种算摇臂被选中的概率的方法，前一种是每选择一次摇臂，就根据公式将所有摇臂的概率更新一次，我觉得这种似乎比较合理，后一种被我注释掉的，是只更新当次被选中的摇臂的概率。两种实验下来结果都差不多，我想问一下到底应该怎么写。

Answer 1

https://blog.csdn.net/pouqiyu5090/article/details/84898609