Why Linear Programming at all?

AI Thread Summary
Linear programming (LP) offers a systematic approach to solving optimization problems with linear inequality constraints, making it easier to identify active constraints compared to using Lagrange multipliers. While Lagrange multipliers can solve these problems, they require manual identification of active constraints, which becomes impractical with larger, real-world problems. The simplex method is highlighted as an efficient algorithm for LP, leveraging the properties of corner points and convex combinations. Additionally, concepts like duals and Kuhn-Tucker conditions extend beyond LP to general nonlinear optimization. Overall, LP is favored for its efficiency and structured methodology in finding solutions.
flyingpig
Messages
2,574
Reaction score
1
I am taking Linear Programming and I haven't completed the course yet, but here is my question.

I've notice that all these problems could have been solved just as easily with Lagrange multipliers.

We got a bunch of linear inequality constraints and an obj function, we can use Lagrange.

Why do we have all of these weird theories about corner points and duals and all?
 
Mathematics news on Phys.org
The basic idea of the LP method is a systematic way to finding which inequality constraints are really equalities (i.e. active constraints) and which are not. Because it is systematic, it is relatively easy to write a computer program to implement it.

If you try to solve the problem using Lagrange multipliers, you have to figure out which constraints are active for yourself, from the Lagrange equations. That might be OK for a problem with 2 or 3 variables and 2 or 3 constraints, but if you try it on a "real world" problem with say 20 variables and 100 constraints (almost all of which are inactive, but you don't know which ones), you need a systematic procedure.

Apart from that, ideas like duals, Kuhn-Tucker conditions, etc also apply to general (nonlinear) optimisation problems, not just to LP problems.
 
Linear programming is one of the successes of Mathematical Optimization (some may say OR). If you look at your Weierstrass theorem and the Local-Global, you can have a very useful result for all LP problems that have a bounded and non-empty constraint set (solution set, also called opportunity set). Thus, what you basically need is an algorithm to find a solution, and by the previous theorems you know is Global.

In terms of applications, many EXIST. You can probably google that. You'll find more than plenty.
 
It is about efficiency. Many areas of mathematics are about solving problems that are easy to solve in principle. There are theoretical and practical reasons for wanting to know if a problem you know can be solved in 10^6 hours with 10^3 computers could be solved in 10^2 hours with 10^1 computers.
 
But we are talking about linear constraints and obj function here. Is it really faster?
 
flyingpig said:
But we are talking about linear constraints and obj function here. Is it really faster?

YES, it is faster because the preference direction (or gradient) of the obj function is constant, and thus the solution for LP is always a CORNER or a convex combination of two corners!. All algorithms look on the boundaries!. In contrast, NonLinear Programs can have a interior solution besides boundary solutions.
 
flyingpig said:
I am taking Linear Programming and I haven't completed the course yet, but here is my question.

I've notice that all these problems could have been solved just as easily with Lagrange multipliers.

We got a bunch of linear inequality constraints and an obj function, we can use Lagrange.

Why do we have all of these weird theories about corner points and duals and all?

The simplex method provides a fast way to solve a linear program on the computer. Do you know of a faster way using Lagrange multipliers?
 

Similar threads

Replies
0
Views
1K
Replies
16
Views
3K
Replies
3
Views
2K
Replies
1
Views
1K
Replies
13
Views
3K
Back
Top