在一个寂寞充实的夜晚,我刷到了这个视频。视频大致的内容是,如果你不在意配偶的颜值,那么择偶时不妨只选丑的,这样你匹配到的配偶在其他你在意的方面会更优秀。这个说法直觉上就很对,但怎么稍微严谨点地证明呢?一开始我想建一个数学模型,但我数学太差了,建不出来,于是就有了这个Python项目和这篇水精华文章。
项目GitHub传送门。作者:hans774882968以及hans774882968以及hans774882968
本文52pojie:https://www.52pojie.cn/thread-1921724-1-1.html
本文juejin:https://juejin.cn/post/7366264344392581146
本文CSDN:https://blog.csdn.net/hans774882968/article/details/138572035
模型
设有N个维度用来衡量男性的优秀程度,组成一个向量(a1, ..., an),0 ,其中a1是颜值,男性的整体分数0 暂且定义为向量各维度的平均数。男性对女性有一个整体分数的要求0 。生成mn_girl_pt的策略:有MALE_LOW_STANDARD_P = 0.75的概率生成小于自身整体分数的数,否则生成一个大于等于自身整体分数的数。
再假设女性的整体分数为0 ,有对男性的整体分数的要求0 ,还有对男性的颜值的期望,用一个函数表示:
AP_THRESHOLD1 = 0.5
AP_THRESHOLD2 = 0.1
e1 = lambda x: x >= AP_THRESHOLD1
e2 = lambda x: x >= AP_THRESHOLD2
# 只选颜值低的
e3 = lambda x: x
其实生成的男性可以改为按分数正态分布,以贴近现实。但有一定难度,就先留个TODO吧。
择偶过程:女性不断地在男性池子中随机抽取男性,若男性和女性互相满足对方的择偶标准,则女性以概率1 - DISCARD_P = 0.5和他成为情侣。女性选择BF_NUM = 5个npy后结束循环。npy中除颜值外的整体分数前SPOUSE_NUM = 1高的为最终的择偶结果,取上述分数的平均值作为最终的择偶分数。保证女性必定能找到配偶。
其实择偶过程可以设计得更复杂些,比如:
[ol]
[/ol]
但这样的机制,相比于让女性必定能找到配偶的机制,要难实现、难分析得多。
择偶时间定义为女性选好BF_NUM个npy时抽取的男性总数。
本文只关注两个指标:匹配到的配偶除颜值外的整体分数,择偶时间。
核心代码讲解
[ol]
[/ol]
女性:
class Female():
def __init__(self, pt: float, mn_boy_pt: float, appearance_jdg) -> None:
self.pt = pt
self.mn_boy_pt = mn_boy_pt
self.appearance_jdg = appearance_jdg
e1 = lambda x: x >= AP_THRESHOLD1
e2 = lambda x: x >= AP_THRESHOLD2
e3 = lambda x: x
我希望同时展示女性自身分数和mn_boy_pt对择偶结果的影响,所以需要画并列的柱状图。所以我将数据结构组织为:
girl_groups = [
[girl1, girl5, girl9], [girl2, girl6, girl10], [girl3, girl7, girl11], [girl4, girl8, girl12]
]
遍历方式:
for j, girl_group in enumerate(girl_groups):
for k, girl in enumerate(girl_group):
boyfriends, time_cost = choose_mate(men, girl)
spouse_pt = calc_spouses_pt(boyfriends)
男性:
class Male():
def __init__(self, key, ap_le=None, ap_geq=None, self_pt_geq=None, girl_pt_le=None) -> None:
self.key = key
pt_vec = get_pt_vector_by_conditions(ap_le, ap_geq, self_pt_geq)
self.appearance = pt_vec[0]
self.other_dimensions = pt_vec[1:]
self.pt = np.mean(pt_vec)
self.pt_without_appearance = np.mean(self.other_dimensions)
self.mn_girl_pt = self.get_mn_girl_pt(girl_pt_le)
def get_mn_girl_pt(self, girl_pt_le=None):
if girl_pt_le is not None:
return random.uniform(0, girl_pt_le)
p = random.random()
if p
key参数仅用于择偶过程去重,其他4个参数都仅用于负责生成满足女性择偶标准的男性的函数ensure_diversity。其实也可以选择不写ensure_diversity,转而编写处理女性没有选到配偶的情况的逻辑。ensure_diversity不太重要但不太好写,放在本节最后介绍。
择偶逻辑:把《模型》一节的择偶过程描述翻译成代码就行,很好写。
def choose_mate(men: List[Male], girl: Female) -> Tuple[List[Male], int]:
BATCH_SIZE = max(1, math.floor(math.sqrt(len(men))))
boyfriends = []
bf_set = set()
time_cost = 0
while True:
men_batch = random.sample(men, BATCH_SIZE)
for man in men_batch:
time_cost += 1
if man.pt girl.pt or not girl.appearance_jdg(man.appearance):
continue
if bf_set.__contains__(man.key):
continue
p = random.random()
if p = BF_NUM:
return boyfriends, time_cost
calc_spouses_pt:计算匹配到的配偶的分数。
def calc_spouses_pt(boyfriends: List[Male]) -> float:
bf_pts = sorted([man.pt_without_appearance for man in boyfriends], reverse=True)
return np.mean(bf_pts[:SPOUSE_NUM])
data_analysis:进行数据分析,画出柱状图。matplotlib画柱状图有点麻烦,我的代码思路参考了参考链接1。ticks表示刻度位置,(-girl_num_in_group / 2 + 0.5) * BAR_WIDTH表示最左侧柱子的左边界位置相比于ticks[?]的偏移。
class ChooseMateResult():
def __init__(self, spouse_pt: float, time_cost: float) -> None:
self.spouse_pt = spouse_pt
self.time_cost = time_cost
def data_analysis(choose_mate_results_groups: List[List[List[ChooseMateResult]]]):
# ...
def auto_label(rects, format_spec):
for rect in rects:
height = rect.get_height() # type(height) = numpy.float64
plt.text(rect.get_x(), 1.01 * height, format(float(height), format_spec), size=8) # 让字小一点,不超过柱子宽度
# 只展示画配偶分数柱状图的代码...
plt.subplot(121)
plt.title('Spouse Points')
girl_num_in_group = len(choose_mate_mean_result_groups[0])
BAR_WIDTH = 0.3
group_num = len(choose_mate_results_groups)
ticks = np.arange(group_num)
girl_in_group_labels = [
f'girls with point {GIRL_PT1}, require {GIRT_REQUIRE_MN_BOY_PT1}',
f'girls with point {GIRL_PT2}, require {GIRT_REQUIRE_MN_BOY_PT1}',
f'girls with point {GIRL_PT2}, require {GIRT_REQUIRE_MN_BOY_PT2}',
]
for j in range(girl_num_in_group):
spouse_pt_mean_bars = [cmg[j].spouse_pt for cmg in choose_mate_mean_result_groups]
x = ticks + (-girl_num_in_group / 2 + 0.5 + j) * BAR_WIDTH
rects = plt.bar(x, spouse_pt_mean_bars, BAR_WIDTH, label=girl_in_group_labels[j])
auto_label(rects, '.3f')
x_labels = [f'girl_group_{i}' for i in range(1, group_num + 1)]
plt.xticks(ticks, x_labels) # 图例
plt.ylim((0, 1)) # 提高 y 范围防止图例遮挡数据
plt.legend()
# ...
ensure_diversity:保证女性能选到配偶的逻辑。
def ensure_diversity(men: List[Male]):
for _ in range(BF_NUM * 2):
ky = len(men)
man1 = Male(ky, ap_geq=AP_THRESHOLD1, self_pt_geq=GIRT_REQUIRE_MN_BOY_PT1, girl_pt_le=GIRL_PT1)
man2 = Male(ky + 1, ap_geq=AP_THRESHOLD2, self_pt_geq=GIRT_REQUIRE_MN_BOY_PT1, girl_pt_le=GIRL_PT1)
man3 = Male(ky + 2, ap_le=AP_THRESHOLD1, self_pt_geq=GIRT_REQUIRE_MN_BOY_PT1, girl_pt_le=GIRL_PT1)
man4 = Male(ky + 3, ap_le=AP_THRESHOLD2, self_pt_geq=GIRT_REQUIRE_MN_BOY_PT1, girl_pt_le=GIRL_PT1)
man5 = Male(ky + 4, ap_geq=AP_THRESHOLD1, self_pt_geq=GIRT_REQUIRE_MN_BOY_PT1, girl_pt_le=GIRL_PT2)
man6 = Male(ky + 5, ap_geq=AP_THRESHOLD2, self_pt_geq=GIRT_REQUIRE_MN_BOY_PT1, girl_pt_le=GIRL_PT2)
man7 = Male(ky + 6, ap_le=AP_THRESHOLD1, self_pt_geq=GIRT_REQUIRE_MN_BOY_PT1, girl_pt_le=GIRL_PT2)
man8 = Male(ky + 7, ap_le=AP_THRESHOLD2, self_pt_geq=GIRT_REQUIRE_MN_BOY_PT1, girl_pt_le=GIRL_PT2)
man9 = Male(ky + 8, ap_geq=AP_THRESHOLD1, self_pt_geq=GIRT_REQUIRE_MN_BOY_PT2, girl_pt_le=GIRL_PT2)
man10 = Male(ky + 9, ap_geq=AP_THRESHOLD2, self_pt_geq=GIRT_REQUIRE_MN_BOY_PT2, girl_pt_le=GIRL_PT2)
man11 = Male(ky + 10, ap_le=AP_THRESHOLD1, self_pt_geq=GIRT_REQUIRE_MN_BOY_PT2, girl_pt_le=GIRL_PT2)
man12 = Male(ky + 11, ap_le=AP_THRESHOLD2, self_pt_geq=GIRT_REQUIRE_MN_BOY_PT2, girl_pt_le=GIRL_PT2)
men.extend((man1, man2, man3, man4, man5, man6, man7, man8, man9, man10, man11, man12))
return men
get_pt_vector_by_conditions不太重要写得也比较丑,不介绍了。只看下它用到的一个我自造的轮子:
def gen_rands_sum_geq_s_range_0_1(n: int, s: float):
if s >= n:
raise ValueError('s is too big. Note: 0 n - i - 1:
legal = False
break
res.append(v)
if not legal or tot
功能是生成n个[0, 1)的数,使得它们的和大于等于s。里面有一个简单的剪枝:s - tot > n - i - 1表示之前生成的数太小,本次生成过程已经失败。TODO: 是否存在一个库能做这件事?
运行结果分析
这里展示了几个比较好的若干个结果。运行拿到比较好的结果还是很容易的,但其实也有不少结果并没有那么好。
输出示意:
[0.532, 0.570, 0.618, 0.621, 0.626, 0.627, 0.656, 0.666, 0.667, 0.677, 0.678, 0.681, 0.709, 0.724, 0.729, 0.730, 0.732, 0.754, 0.770, 0.789] 0.678 44.2
[0.568, 0.605, 0.631, 0.634, 0.652, 0.667, 0.671, 0.680, 0.687, 0.701, 0.704, 0.711, 0.719, 0.732, 0.737, 0.737, 0.738, 0.773, 0.800, 0.823] 0.698 40.6
[0.611, 0.628, 0.648, 0.653, 0.658, 0.701, 0.708, 0.711, 0.719, 0.724, 0.728, 0.734, 0.736, 0.748, 0.748, 0.761, 0.775, 0.789, 0.813, 0.884] 0.724 84.2
[0.559, 0.560, 0.572, 0.610, 0.616, 0.620, 0.624, 0.626, 0.627, 0.676, 0.680, 0.682, 0.694, 0.716, 0.732, 0.738, 0.749, 0.772, 0.831, 0.831] 0.676 32.2
[0.549, 0.613, 0.643, 0.662, 0.664, 0.671, 0.679, 0.685, 0.687, 0.687, 0.700, 0.732, 0.738, 0.743, 0.749, 0.762, 0.777, 0.792, 0.794, 0.868] 0.710 32.0
[0.659, 0.683, 0.697, 0.704, 0.724, 0.725, 0.730, 0.731, 0.756, 0.764, 0.777, 0.781, 0.787, 0.792, 0.797, 0.829, 0.850, 0.859, 0.864, 0.891] 0.770 63.2
[0.663, 0.679, 0.688, 0.694, 0.706, 0.706, 0.709, 0.725, 0.733, 0.751, 0.756, 0.767, 0.774, 0.781, 0.781, 0.786, 0.787, 0.812, 0.817, 0.865] 0.749 95.6
[0.604, 0.604, 0.644, 0.645, 0.659, 0.669, 0.673, 0.687, 0.706, 0.729, 0.739, 0.744, 0.748, 0.760, 0.766, 0.768, 0.775, 0.825, 0.862, 0.865] 0.723 80.0
[0.683, 0.722, 0.740, 0.745, 0.750, 0.759, 0.762, 0.770, 0.787, 0.794, 0.800, 0.819, 0.819, 0.828, 0.854, 0.868, 0.871, 0.883, 0.888, 0.924] 0.803 242.2
[0.681, 0.699, 0.746, 0.757, 0.758, 0.762, 0.768, 0.776, 0.785, 0.790, 0.807, 0.815, 0.819, 0.829, 0.831, 0.833, 0.839, 0.840, 0.888, 0.888] 0.795 628.8
[0.639, 0.680, 0.705, 0.708, 0.712, 0.719, 0.720, 0.724, 0.735, 0.746, 0.759, 0.760, 0.760, 0.770, 0.807, 0.820, 0.836, 0.841, 0.842, 0.862] 0.757 458.8
[0.775, 0.787, 0.795, 0.797, 0.812, 0.816, 0.819, 0.827, 0.832, 0.840, 0.845, 0.845, 0.846, 0.850, 0.875, 0.879, 0.888, 0.896, 0.904, 0.904] 0.842 2397.6
柱状图示意:
choose_ugly_mates4.png (52.75 KB, 下载次数: 0)
下载附件
2024-5-8 14:40 上传
描述一下图中的信息:
[ol]
[/ol]
结论(TLDR)
一、只提高自身分数,不一定能提升匹配到的配偶的分数,提升的点大概率体现在择偶时间的缩短。
二、提高择偶标准,可以有效提升匹配到的配偶的分数,但会增加择偶时间,且可能增加到无穷。
三、若其他条件不变,对颜值的期望分别如下:
AP_THRESHOLD1 = 0.5
AP_THRESHOLD2 = 0.1
e1 = lambda x: x >= AP_THRESHOLD1
e2 = lambda x: x >= AP_THRESHOLD2
# 只选颜值低的
e3 = lambda x: x
则大概率匹配到的配偶的除颜值外的整体分数满足e1 。如果分数不单增,那么大概率提升的点体现在择偶时间的缩短。
四、只考虑选颜值低的为配偶的策略确实能提升除颜值外的整体分数,但也会显著增加择偶时间。具体牺牲哪个,优化哪个,看你自己咯。
参考资料
[ol]
[/ol]