sklearn集成学习 关于 算法准确率的两个问题

Question

在学习sklearn中的集成学习中遇到了两个问题：

1. 集成学习中soft voting的准确率低于hard voting。
   代码如下：

import numpy as np
import matplotlib.pyplot as plt
from sklearn import datasets
x,y = datasets.make_moons(n_samples = 500,noise = 0.3, random_state = 42)
from sklearn.model_selection import train_test_split
x_train,x_test,y_train,y_test = train_test_split(x,y,random_state = 42)
from sklearn.ensemble import VotingClassifier
voting_clf = VotingClassifier(estimators = [
    ('log_clf',LogisticRegression()),
    ('svm_clf',SVC()),
    ('dt_clf',DecisionTreeClassifier())
],voting = 'hard')
voting_clf.fit(x_train,y_train)
voting_clf.score(x_test,y_test)
voting_clf2 = VotingClassifier(estimators = [
    ('log_clf',LogisticRegression()),
    ('svm_clf',SVC(probability = True)), #修改SVC参数
    ('dt_clf',DecisionTreeClassifier(random_state = 666))],voting = 'soft')
voting_clf2.fit(x_train,y_train)
voting_clf2.score(x_test,y_test)# soft 与hard 的结果都是0.904 很奇怪

1. 集成学习中采用决策树的数量增多，准确率并没有提高。
   代码如下：

   import numpy as np
   import matplotlib.pyplot as plt
   from sklearn import datasets
   x,y = datasets.make_moons(n_samples = 500,noise = 0.3, random_state = 42)
   from sklearn.model_selection import train_test_split
   x_train,x_test,y_train,y_test = train_test_split(x,y,random_state = 42)
   bagging_clf = BaggingClassifier(DecisionTreeClassifier(),
                               n_estimators =500, max_samples = 100,bootstrap = True)
   # n_estimator 多少个子模型 max_samples看多少样本 bootstrap是否放回
   %%time
   bagging_clf.fit(x_train,y_train)# 500个决策树
   bagging_clf.score(x_test,y_test)
   single_dec_tree = DecisionTreeClassifier()
   single_dec_tree.fit(x_train,y_train)# 1个决策树
   single_dec_tree.score(x_test,y_test)
   bagging_clf5000 = BaggingClassifier(DecisionTreeClassifier(),
                                   n_estimators = 5000,max_samples = 100,bootstrap = True)
   # 5000个决策树
   %%time
   bagging_clf5000.fit(x_train,y_train)
   bagging_clf5000.score(x_test,y_test)
   # 单个决策树的准确率为0.88，500个是0.928，5000个是0.912
   

这两个问题类似，因为从算法的原理上讲，soft的结果至少不会比hard的差，为啥结果会这样呢？
第二个问题中，子模型的数目增多，一定会使得整体的准确率提高，但是为啥决策树的数目增多，准确率不升反降？
难道存在一些子模型的准确率低于平均准确率吗？很奇怪