本博客基本是ceres 的翻译,更多信息请查看 http://ceres-solver.org/nnls_solving.html
待求解最小二乘问题:
其中: and
求解方法主要分为两类:
1. Trust Region methods
2. linear search methods
1. Trust Region Methods
算法步骤:
其中 是trust region radius,
is a measure for the the performance that predicts the behavior of the non-linear objective
求解上述问题主要有两种方法:Levenberg-Marquardt and Dogleg
a. LM methods
b. Dogleg methods
2. Line Search Methods
算法步骤
Here is some approximation to the Hessian of the objective function, and
is the gradient at
. Depending on the choice of
we get a variety of different search directions
所以:
1. trust region 和 line search 方法是相反的两种方法,trust region 方法先确定半径,然后再确定move dirction. line search methd 先确定dirction , 然后确定步长
2. line search method 只能处理不含约束的最小二乘问题,但是trust region 方法可以处理含约束的最小二乘问题,但是含约束的也可以通过添加松弛变量的方式转成不含约束问题。