zoukankan      html  css  js  c++  java
  • 凸优化和对偶

    一、机器学习基础及凸优化

    参考:http://cvxopt.org/userguide/coneprog.html

    1. 凸函数

    1.1 Optimization Categories

    1.1.1 convex or non-convex

    lGlobal optimization or better local optimization

    lconvex set:假设对于任意x,y∈C并且任意参数,a∈[0,1],我们对ax+(1-a)y∈C   https://zhuanlan.zhihu.com/p/92230334

    lConvex Function define:函数的定义域domf为凸集,对于定义域里任意x,y,函数满足f(θx + (1-θy))<=θf(x)+(1-θ)f(y)   
     https://www.zhihu.com/question/20014186/answer/27194360

    1.1.2 continuous or discrete

    1.1.3 constraint or non-constraint

    1.1.4 smooth or non-smooth

    1.2 问题解决过程:

    lDecision Variable

    lObjective Function

    lConstraint

    l判断类型

    l设计或使用

    1.3 应用

    lLP:Transportation(运输) Problem: min Transportation cost  minf s.t.  条件

    lportfolio optimization:10万块钱-->买多支股票    Mean Variance portfolio optimization

    lset cover problem:找最少集合的个数

    lExhaustive Search :枚举(NP-hard的时候可用)

    lGreedy search:Local method-->global optimization

    lnon-convex --> relax -->convex

    2. duality(对偶):视角不同-->minimize primal and maximize dual(见图,理想情况下会相遇)  凹函数

     lprimal-->dual

    lLower bound property:P*>=d*

    lStrong and weak Duality:结果可能不一样

    lstrong条件:Conplementary Slackness

    2.1 strong条件:KKT conditions

     

     

    一、机器学习基础及凸优化

    参考:http://cvxopt.org/userguide/coneprog.html

    1. 凸函数

    1.1 Optimization Categories

    1.1.1 convex or non-convex

    Global optimization or better local optimization

    convex set:假设对于任意x,y∈C并且任意参数,a∈[0,1],我们对ax+(1-a)y∈C   https://zhuanlan.zhihu.com/p/92230334

    Convex Function define:函数的定义域domf为凸集,对于定义域里任意x,y,函数满足f(θx + (1-θy))<=θf(x)+(1-θ)f(y)   
     https://www.zhihu.com/question/20014186/answer/27194360

    1.1.2 continuous or discrete

    1.1.3 constraint or non-constraint

    1.1.4 smooth or non-smooth

    1.2 问题解决过程:

    Decision Variable

    Objective Function

    Constraint

    判断类型

    设计或使用

    1.3 应用

    LP:Transportation(运输) Problem: min Transportation cost  minf s.t.  条件

    portfolio optimization:10万块钱-->买多支股票    Mean Variance portfolio optimization

    set cover problem:找最少集合的个数

    Exhaustive Search :枚举(NP-hard的时候可用)

    Greedy search:Local method-->global optimization

    non-convex --> relax -->convex

     

    2. duality(对偶):视角不同-->minimize primal and maximize dual(见图,理想情况下会相遇)  凹函数

     

    primal-->dual

     

     

    Lower bound property:P*>=d*

    Strong and weak Duality:结果可能不一样

    strong条件:Conplementary Slackness

      

    2.1 strong条件:KKT conditions

      

      

  • 相关阅读:
    RNN-2-前向传播、BPTT反向传播
    RNN-1-参数共享、网络的展开、常见应用
    被围绕的区域
    语言模型的评价方法
    个性化推荐系统
    推荐系统-CTR-Wide&Deep
    推荐系统-CTR-总结
    推荐系统-CTR-PNN
    Local variable flag defined in an enclosing scope must be final or effective
    2.两数相加
  • 原文地址:https://www.cnblogs.com/Towerb/p/14012244.html
Copyright © 2011-2022 走看看