Proximal Gradient
是一种一般化的解决 non-differentiable convex optimization
的方法. 即如下:
其中所有f都是convex的,但部分是不可求导的
,这样就不能用传统的Gradient Descent方法。
Proximal gradient就是用来解决这类问题的,其主要分开求解,即每个函数独自的考虑(They a re called proximal because each non smooth function among f_1, . . . , f_n is involved via its proximity operator), 比较典型的这类算法有: ISTA、, projected Landweber, projected gradient, alternating projections等,详细讨论见[1]
Referance
- Combettes. “Proximal Splitting Methods in Signal Processing”