關(guān)于learning feasible和hoeffding bound的理解，該實(shí)驗(yàn)很有意思：

image.png

(a) u=0.5
(b) 樣本是抽一枚硬幣，樣本集D是抽10次，總共100,000次試驗(yàn)，相當(dāng)于生成了100,000個(gè)樣本集，這里面有g(shù)ood data, 當(dāng)然也有bad data. 代碼如下：

from random import randint
from matplotlib import pyplot as plt
import os
import json
import collections

import numpy as np


class Coin:
    """1是head"""

    def roll(self):
        return randint(0, 1)


class Experiment:

    def __init__(self):
        self.coins = []
        self.result = []
        self._init_coins()

    def _init_coins(self):
        for _ in range(1000):
            self.coins.append(Coin())

    def execute_experiment(self):
        for coin in self.coins:
            total = 0
            for _ in range(10):
                total += coin.roll()
            self.result.append(total / 10)
        v1 = self.result[0]
        v_rand = self.result[randint(0, 999)]
        v_min = min(self.result)
        return v1, v_rand, v_min


def load_data():
    with open('v1.json', 'r') as f:
        v1_data = json.load(f)
    with open('v_rand.json', 'r') as f:
        v_rand_data = json.load(f)
    with open('v_min.json', 'r') as f:
        v_min_data = json.load(f)
    print_counts(v1_data)
    print_counts(v_rand_data)
    print_counts(v_min_data)
    plt.subplot(131)
    plt.hist(v1_data, bins=11)
    plt.subplot(132)
    plt.hist(v_rand_data, bins=11)
    plt.subplot(133)
    plt.hist(v_min_data, range=(0, 1))
    plt.show()
    return v1_data, v_rand_data, v_min_data


def print_counts(l):
    obj = collections.Counter(l)
    for k, v in sorted(obj.items()):
        print(k, v)
    print()

(c) v已經(jīng)算出來(lái)，u=0.5, 把epsilon從0到1逐漸增大，概率P可以用頻率來(lái)得到，繪圖函數(shù)時(shí)plot_bound

def plot_bound(l_v, legend):
    np_v = np.array(l_v)
    total_experiments = len(l_v)
    abs_diff = np.abs(np_v - 0.5)  # u=0.5
    N = 10
    P = []
    upper_bounds = []
    epsilons = np.linspace(0, 1, 50, endpoint=False)
    for epsilon in epsilons:
        P.append(sum(abs_diff > epsilon) / total_experiments)
        upper_bounds.append(2 * np.exp(-2 * epsilon ** 2 * N))
    plt.plot(epsilons, P)
    plt.plot(epsilons, upper_bounds)
    plt.legend([legend, 'upper bound'], loc='upper right')
    plt.show()


if __name__ == '__main__':
    if not os.path.exists('v1.json'):
        v1_data = []
        v_rand_data = []
        v_min_data = []
        for _ in range(100000):
            experiment = Experiment()
            v1, v_rand, v_min = experiment.execute_experiment()
            v1_data.append(v1)
            v_rand_data.append(v_rand)
            v_min_data.append(v_min)

        with open('v1.json', 'w') as f:
            json.dump(v1_data, f)
        with open('v_rand.json', 'w') as f:
            json.dump(v_rand_data, f)
        with open('v_min.json', 'w') as f:
            json.dump(v_min_data, f)
    v1_data, v_rand_data, v_min_data = load_data()
    plot_bound(v1_data, 'v1')
    plot_bound(v_rand_data, 'v_rand')
    plot_bound(v_min_data, 'v_min')

繪出下面3幅圖：

image.png

image.png

image.png

image.png