论文解读----A SURVEY ON MULTIOBJECTIVE DESCENT METHODS

针对文章中的Section 4中带约束的优化问题进行解读.

问题

minF(x),\min F(x),

s.t.xCs.t. \quad x\in C

的pareto efficient的必要条件是(J表示Jacobian Matrix):

JF(xˉ)(Rn)[R++m]=JF(\bar{x})(\mathbb{R}^n) \cap [-\mathbb{R}_{++}^m] = \emptyset

where

JF(xˉ)(Rn):={JF(xˉ)v,vRn}JF(\bar{x})(\mathbb{R}^n) := \{JF(\bar{x})v, v\in \mathbb{R}^n \}

那么,对于scalar-valued optimization其必要条件为:

<F(xˉ),xxˉ>0<\nabla F(\bar{x}), x - \bar{x}> \geq 0

定义带有L2L_2正则项的梯度搜索方向为(得到的结果为下一个x的point位置,与当前x相减组成的向量方向):

v(x):=arg minvCx(βφx(v)+12v2)v(x):= \argmin_{v\in C-x} (\beta \varphi_x(v) + \frac{1}{2} \lVert v \rVert^2)

v(x)v(x)是强凸且well defined, 定义最优的value:

θ(x):=minvCx(βφx(v)+12v2)=(βφx(v(x))+12v(x)2)\theta(x):= \min_{v\in C -x}(\beta \varphi_x(v) + \frac{1}{2} \lVert v \rVert ^2) = (\beta \varphi_x(v(x)) + \frac{1}{2} \lVert v(x) \rVert ^2)

Proposition 4.1.
xx is stationary if and only if θ(x)=0\theta(x)=0
proof: