当前位置:网站首页>数学建模——红酒品质分类
数学建模——红酒品质分类
2022-07-29 01:08:00 【abcwsp】
贝叶斯优化是一种基于概率模型下的超参数调优方式。通过不断地添加样本点来更新目标函数的后验分布,从而更好的调整当前参数。由于贝叶斯优化在低维度下迭代次数少,速度快,对非凸问题稳健,因此常用于对提升树的参数优化。贝叶斯优化包含高斯过程回归和贝叶斯开发探索两个过程。
高斯过程回归
相比于传统获取后验分布模型,高斯过程考虑到 y N y_N yN与 y N + 1 y_{N+1} yN+1的关系即:
p ( y N + 1 ∣ X N + 1 , y N ) \begin{align} p\left(y_{N+1} \mid X_{N+1}, y_{N}\right) \end{align} p(yN+1∣XN+1,yN)
高斯过程通过假设 Y Y Y服从联合正态分布来考虑 y N y_N yN和 y N + 1 y_{N+1} yN+1之间的关系:
[ y 1 y 2 … y n ] ∼ N ( 0 , [ k ( x 1 , x 1 ) , k ( x 1 , x 2 ) , … , k ( x 1 , x n ) k ( x 2 , x 1 ) , k ( x 2 , x 2 ) , … , k ( x 2 , x n ) … k ( x n , x 1 ) , k ( x n , x 2 ) , … , k ( x n , x n ) ] ) \begin{align} \left[\begin{array}{c} y_{1} \\ y_{2} \\ \ldots \\ y_{n} \end{array}\right] \sim N\left(\mathbf{0},\left[\begin{array}{c} k\left(x_{1}, x_{1}\right), k\left(x_{1}, x_{2}\right), \ldots, k\left(x_{1}, x_{n}\right) \\ k\left(x_{2}, x_{1}\right), k\left(x_{2}, x_{2}\right), \ldots, k\left(x_{2}, x_{n}\right) \\ \ldots \\ k\left(x_{n}, x_{1}\right), k\left(x_{n}, x_{2}\right), \ldots, k\left(x_{n}, x_{n}\right) \end{array}\right]\right) \end{align} ⎣⎡y1y2…yn⎦⎤∼N⎝⎛0,⎣⎡k(x1,x1),k(x1,x2),…,k(x1,xn)k(x2,x1),k(x2,x2),…,k(x2,xn)…k(xn,x1),k(xn,x2),…,k(xn,xn)⎦⎤⎠⎞
其中协方差矩阵为核矩阵,仅与特征 x x x有关。将 Y Y Y服从高维高斯分布作为先验条件,根据训练集得到最优核矩阵,从而得到后验分布:
p ( y ∗ ∣ y ∼ N ( K ∗ K − 1 y , K ∗ ∗ − K ∗ K − 1 K ∗ T ) \begin{align} p\left(y_{*} \mid \mathbf{y} \sim N\left(K_{*} K^{-1} \mathbf{y}, K_{* *}-K_{*} K^{-1} K_{*}^{T}\right)\right. \end{align} p(y∗∣y∼N(K∗K−1y,K∗∗−K∗K−1K∗T)
贝叶斯优化
通过高斯过程回归对目标函数建模,得到其后验分布后,需要抽样进行样本计算,而贝叶斯优化很容易在局部最优解上不断采样,为了保证开发和探索之间的权衡。需要使用Acquisition Function寻求下一个 x x x的函数。其中POI函数是一种主流的选择方式:
POI ( X ) = P ( f ( X ) ≥ f ( X + ) + ξ ) = Φ ( μ ( x ) − f ( X + ) − ξ σ ( x ) ) \operatorname{POI}(X)=P\left(f(X) \geq f\left(X^{+}\right)+\xi\right)=\Phi\left(\frac{\mu(x)-f\left(X^{+}\right)-\xi}{\sigma(x)}\right) POI(X)=P(f(X)≥f(X+)+ξ)=Φ(σ(x)μ(x)−f(X+)−ξ)
其中 f ( X ) f(X) f(X)为 X X X的目标函数值, f ( X + ) f(X^{+}) f(X+)为当前目标函数最优值, u ( x ) , σ ( x ) u\left( x \right) ,\sigma \left( x \right) u(x),σ(x)分别为高斯过程所得到的目标函数的均值和方差。贝叶斯优化即不断尝试新的 X X X使得 P O I ( X ) POI(X) POI(X)最大

边栏推荐
- Leetcode 113: path sum II
- About df['a column name'] [serial number]
- HCIA configuration instance (ENSP)
- 【Golang】- runtime.Goexit()
- [WesternCTF2018]shrine
- We summarized the three recommendations for the use of Nacos and first published the Nacos 3.0 plan for the 4th anniversary of the open source of Nacos
- [7.21-26] code source - [good sequence] [social circle] [namonamo]
- Analysys analysis: focus on users, improve the user experience of mobile banking, and help the growth of user value
- Data security is a competitive advantage. How can companies give priority to information security and compliance
- 【公开课预告】:快手GPU/FPGA/ASIC异构平台的应用探索
猜你喜欢

Thirty years of MPEG audio coding

560 and K

T-sne dimensionality reduction

给LaTeX公式添加优美的注解;日更『数据科学』面试题集锦;大学生『计算机』自学指南;个人防火墙;前沿资料/论文 | ShowMeAI资讯日报

The information security and Standardization Commission issued the draft for comments on the management guide for app personal information processing activities

Six noteworthy cloud security trends in 2022

ASCII code table

Sigma-DSP-OUTPUT

JS 定时器setInterval clearInterval 延时器setTimeOut 异步 动画

Timer of BOM series
随机推荐
Overview of Qualcomm 5g intelligent platform
【流放之路-第八章】
Leetcode 112: path sum
DSP vibration seat
【7.21-26】代码源 - 【平方计数】【字典序最小】【“Z”型矩阵】
抓包工具Charles使用
[hcip] experiment of republishing and routing strategy
Analyzing the function of human-computer interface module of runtime manager based on autoware
Data security is a competitive advantage. How can companies give priority to information security and compliance
ELS stop at all
Thirty years of MPEG audio coding
What are the common cyber threats faced by manufacturers and how do they protect themselves
Network security litigation risk: four issues that chief information security officers are most concerned about
【流放之路-第七章】
Reinforcement learning (III): dqn, nature dqn, double dqn, with source code interpretation
【7.27】代码源 - 【删数】【括号序列】【数字替换】【游戏】【画画】
Focus on differentiated product design, intelligent technology efficiency improvement and literacy education around new citizen Finance
Secret skill winter tide branding skill matching
Introduction to Elmo, Bert and GPT
T-sne dimensionality reduction