Title: ESI 4313 Operations Research 2
1ESI 4313Operations Research 2
- Nonlinear Programming Model
- (Lecture sets 45)
2Differences betweenLP and NLP
- The feasible region of an LP is always a convex
set. - in fact, a polyhedron
- Isocost/profit curves are straight lines (or
planes) - an LP always has an extreme point optimal
solution (if an optimal solution exists)
3Finding the optimal solution for an LP
x2
20x110x2160
10
x16
z40000
(4½,7)
z20000
10x115x2150
5
z50500
x1
8
6
4
4Differences betweenLP and NLP
- These results do not generalize to NLP
- Consider the following example
- A company can produce a good using capital and
labor. - K units of capital and L units of labor yields KL
units of the good - Capital can be purchased at 200/unit, while
labor can be purchased at 1/unit - A total budget of 40 is available
5Example 1
- NLP model
- maximize z KL
- subject to
- 200K L ? 40
- K, L ? 0
6Example 1
- Ignoring the budget constraint, we can draw the
isoprofit lines for the objective function - Rewrite z KL as L z/K
- Plot this function for different values of z
7Isoprofit Lines
8Example 1
- Now also plot the feasible region of the NLP
problem
L 40 200 K
9Differences betweenLP and NLP
- We see that for an NLP the optimal solution is
not necessarily an extreme point of the feasible
region even if the feasible region is a
polyhedron - But in the example the optimal solution is on
the boundary of the feasible region - Is this true in general?
10Example 2
- Consider a simplified version of the location
problem with Euclidean distances - In particular, let there be only 1 customer
location, (2,2) - But suppose that the facility location is
constrained to be in a particular area
11Example 2a
- Suppose that we require that the facility is in
the square 0,1?0,1, i.e., we have the
constraints - so the full problem is
12Example 2a
? y
(2,2)
x ?
13Example 2a
- Again, the optimal solution is on the boundary of
the feasible region - Again, is this true in general?
14Example 2b
- Suppose that we require that the facility is in
the square 0,3?0,3, i.e., we have the
constraints - so the full problem is
15Example 2b
? y
(2,2)
x ?
16Example 2b
- In this case, the optimal solution is in the
interior of the feasible region - Conclusion
- For general NLP problems, we have lost the
structure of optimal solutions that we found for
LP problems - We need different approaches, concepts, etc. to
help us solve such problems
17Example 3
- We could have found an even simpler example
- minimize f(x) x2
- subject to
- -1 ? x ? 1
18One-dimensional NLP problems
- A problem with only a single decision variable is
called a one-dimensional optimization problem. - If we restrict ourselves to LP, one-dimensional
problems are trivially solvable. (why? how?) - But even one-dimensional NLP problems can be
challenging!
19Example 4
- minimize f(x)
- subject to
- 0 ? x ? 360
20Local Optima
- For any NLP problem, a feasible vector x is a
- local maximum if f(x) ? f(x) for all (feasible)
x that are close to x. - local minimum if f(x) ? f(x) for all (feasible)
x that are close to x.
21Local Optima
- What does for all x that are close to x mean?
- For all x such that
22Example 4
examples of local maximum
examples of local minimum
23Local Optima
- Many algorithms exist for finding a local optimum
of an NLP - some of these we will study in this class
- Solvers like LINGO and Excel Solver only find a
local optimum !! - if you want to find the global optimum, you need
to do more work
24Example 2
Local optimal solution found at iteration
22 Objective value
5058.530
Variable Value Reduced Cost
X 3.200526
0.000000 Y
6.220236 0.000000
Row Slack or Surplus Dual
Price 1
5058.530 -1.000000
25Local Optima
- For an LP problem, any local optimum is a global
optimum - why?
- Can we find types of NLP problems for which this
also holds? - this would ensure that LINGO or Excel finds the
global optimum of the NLP
26Convex and concave functions
- A very important class of nonlinear functions are
convex and concave functions - Let f(x1,,xn) be a function defined for all
vectors x(x1,,xn) in some convex set S
27Convex sets
convex
non-convex
28Convex sets
- Recall that a set is convex if the line segment
joining any pair of points in the set is
completely contained in the set - More formally
29Convex functions
- The function f is a convex function if
30Example convex function
?x(1- ?)x
x
x
31Concave functions
- The function f is a concave function if
32Example concave function
x
x
?x(1- ?)x
33Examplenonconvex/nonconcave function
x
x
34Importance of concave functions in NLP
- Suppose that we have an NLP with the following
properties - the feasible region, say S, is convex
- the objective function, say f, is concave
- the objective is to maximize the value of the
objective function
35Importance of concave functions in NLP
- Then
- Any local maximum is a global maximum!
- How can we prove this mathematically?
- using a proof by contradiction
- we assume that the result is false, and then
derive a contradiction, i.e., something that is
clearly not true - this then implies that the result must be true
36Importance of concave functions in NLP
- Suppose there exists a solution x that is a
local maximum, but not a global maximum - Since x is not a global maximum, there exists a
solution x with the property that f(x) gt f(x) - Now use the fact that f is concave
37Importance of concave functions in NLP
- Since f is concave, we have that
38Importance of concave functions in NLP
- If ? is very close to 1, then
- ?x(1-?)x is very close to x
- f (?x(1-?)x ) gt f (x )
- Therefore, x cannot be a local optimum
- This contradicts our assumption that there exists
a local maximum that is not a global maximum - Thus the result is true!!
39Importance of convex functions in NLP
- Suppose that we have an NLP with the following
properties - the feasible region, say S, is convex
- the objective function, say f, is convex
- the objective is to minimize the value of the
objective function
40Importance of convex functions in NLP
- Then
- Any local minimum is a global minimum!
- This can be proven mathematically in a similar way