Like all constrained optimization (mathematical programming) problems, linear programs have three basic ingredients: decision variables, which are items whose quantities that we can choose or control; an objective function, whose value is a function of the decision variables and is minimized or maximized; and constraints (, =, ) , which are also defined in terms of the decision variables and must be satisfied for the solution to the problem to be acceptable. More specifically, a constraint can always be formulated so that its left-hand-side (the stuff to the left of the , =, or ) is a function of the decision variables, and the right-hand-side is a constant, a simple number, rather than a formula.
In constrained optimization problems, a solution refers to choice of a specific set of values for the decision variables, while the Objective Function Value (OFV) is the value the objective function takes on for a solution. For example. if the objective function is defined to be
2 X 1 +6 X 2 ,
then the solution (3 , 2) produces an OFV of 18. A feasible solution is a choice of values that satisfies all of the constraints, while an infeasible solution is a choice for which at least one constraint is violated. An optimal solution is a feasible solution for which the OFV cannot be improved upon -- further minimized or maximized -- by any other feasible solution. Note that there may be more than one optimal solution.
Unlike other constrained optimization models, LPs always have a linear objective function and linear functions on the left-hand-sides of their constraints. Mathematically, a linear function of the decision variables X 1 , X 2 , . . ., X n can always be written as
f(X 1 , X 2 , . . ., X n ) =a 1 X 1 +a 2 X 2 +. . . +a n X n ,
a function for which each decision variable is multiplied by a coefficient and added to its peers. Two examples of nonlinear functions are
f(X 1 , X 2 ) =a X 1 X 2 and f(X 1 , X 2 , X 3 ) =a 1 X 1 +a 2 X 2 +a 3 log(X 3 ).
Practically, linearity is intended to model situations in which the quantities represented by the decision variables are divisible and for which the function changes proportionally with changes in each variable. Continuous quantities, such as tons of steel to produce and gallons of gasoline to consume, are divisible, but lumpy quantities, such as numbers of aircraft carriers to produce, Background on Constrained Optimization and LP Decision Models & Optimization Page 2 are not. That is, if we constructed and solved an LP whose optimal solution included the consumption of 5.57 gallons of gasoline, the answer would make sense to us; we are able to consume fractions of gallons of gasoline. If the optimal solution called for the production of 5.57 aircraft carriers, however, the solution probably wouldnt make sense, since 0.57 aircraft carriers arent of much use to anyone. Even if we assume the value is 0, its difficult to describe how to measure fractional aircraft carriers, or what the costs of producing them would be. The proportionality idea is similar: while the percentage of fat in dog food changes in proportion to the mix and fat-content of its ingredients, the area of a circle ( r 2 ) is not proportional to its radius or circumference.
Later in the course, we will discuss how to use techniques similar to linear programming to allow us to relax the assumptions of divisibility and proportionality.
One final assumption implicit in the formulation of LPs (and of constrained optimization problems, in general) is that the collection of numbers that make up the objective function coefficients, constraint coefficients, and right-hand-sides are fixed quantities that are known with certainty. In real life, however, there is often uncertainty about the values these data take on, and we are forced to use estimates of what the numbers are. For example, in the perfume production problem, we might not know with certainty the contribution from selling each ounce of perfume, because we anticipate industry-wide price increases to take effect in the next selling period. It would still be useful, however, to run the LP model developed for the problem, using our best estimate of what the contribution will bewith the understanding that we may need to take the models recommendation with a grain of salt when making our management decision.