b****t
发帖数: 114 | 1 Hello all,
I am thinking about optimizing a piecewise linear function. Since the
explicit function form is too complicated, I only know the function form
piecewisely, i.e. a linear function for a small subset of the domain (e.g.
points in a unit d-dimensional simplex). Thus the overall function is
piecewise linear over the R^d space with integer breaking points. ( I hope I
state the problem clearly here).
The steepest descent method can be used to find the optimum of the function
here? Dose the | D*******a
发帖数: 3688 | 2 你的objective func是convex么
I
function
way
【在 b****t 的大作中提到】
: Hello all,
: I am thinking about optimizing a piecewise linear function. Since the
: explicit function form is too complicated, I only know the function form
: piecewisely, i.e. a linear function for a small subset of the domain (e.g.
: points in a unit d-dimensional simplex). Thus the overall function is
: piecewise linear over the R^d space with integer breaking points. ( I hope I
: state the problem clearly here).
: The steepest descent method can be used to find the optimum of the function
: here? Dose the
| b****t
发帖数: 114 | 3
Hi DrumMania,
I do not have any convexity assumption on the objective function. If its
convex, then I think subgradient method and guarantee the convergence of
search to the optimum.
Can I ask the question another way: the steepest descent search method can
guarantee the convergence for what type of obj functions? Continuous and
smooth, and unimodular?
Thanks again,
Beet
【在 D*******a 的大作中提到】
: 你的objective func是convex么
:
: I
: function
: way
| D*******a
发帖数: 3688 | 4 步长设置得当的话肯定能收敛到一个local minimum，只是不能判定是否global
【在 b****t 的大作中提到】
:
: Hi DrumMania,
: I do not have any convexity assumption on the objective function. If its
: convex, then I think subgradient method and guarantee the convergence of
: search to the optimum.
: Can I ask the question another way: the steepest descent search method can
: guarantee the convergence for what type of obj functions? Continuous and
: smooth, and unimodular?
: Thanks again,
: Beet
|
|
