Simple but not simple Nonconvex Optimization Problem

132 Views Asked by At

everyone. Just bump into one nonconvex optimization problem. Looks simple, but I dont know how to solve it. The problem is

$\max_{a,b,\tau} \tau$

$\text{s.t.} \tau\leq \tau_{1}(a,b),$

$\tau\leq \tau_{2}(a,b),$

$a_{\text{min}}\leq a \leq a_{\text{max}},$

$b_{\text{min}}\leq b \leq b_{\text{max}},$

where $\tau_{1}(a,b)$ and $\tau_{2}(a,b)$ are not convex functions in terms of either $a$ or $b$.

Initially, I think maybe I can make use of duality and get its Lagrangian dual problem. Since the dual problem is always convex, I can use some off-the-shelf algorithms to solve dual problem. But I realize that what I can find is still a local optimal solution, not the global one. I know I can work on something on $\tau_{1}$ and $\tau_{2}$ (e.g., whether they are monotonically increasing functions of $a$ or $b$), but I am wondering is there any general way to solve such nonconvex problem? Thank you.

1

There are 1 best solutions below

2
On BEST ANSWER

First of all, it is hard to understand your main question. 'Simple but not simple' is not reflecting your wish to solve a nonconvex problem. I assume your variables are continuous, e.g. $\tau \in \mathbb{R}$, so that the last two constraints will not be combinatorial constraints. I conclude they are just continuous simple constraints.

You have non-convex constraints, $\tau \leq \tau_i(a,b) $ which are the main problems. You can end up at a local minimum, be careful about it. And you are planning to use Lagrangian duality in a non-convex problem. That's another issue. You can just get a bound on the optimal value, where strong duality will most probably not hold.

What I suggest is you to write more details about the functions $\tau_i$ and try to analytically derive something. Search for convex hulls, SDP relaxations or whatever is relevant to your functions.

If you want to solve a specific problem, instead of analytically deriving a solution, you can always try to solve in Artelys Knitro solver. It is doing pretty well in dirty non-linear models.