Operations Research Lecture 1: Linear Programming Introduction

Transcription

1 Operations Research Lecture 1: Linear Programming Introduction Notes taken by Kaiquan School, Nanjing University 25 Feb Some Real Problems Some problems we may meet in practice or academy: 1.1 Production Planning Given a manufacturer plans to produce two types of products: I and II, the required matireials (A & B) and equipment for producting one product are list in Table 1. The profits for Product I and II are 2$ and 3$. The question is: how to make product planning, so the manufacturer can get the maximum profit. Table 1: Example: Product Planning. Product I Product II Available Resources equipment matireial A matireial B Let x 1 and x 2 be as the numbers of Product I and II to be produced. Under the constraints of the resources, the variables should satisfy the following conditions: x 1 + 2x 2 8 4x x 2 12 The profit z can be represented as z = 2x 1 + 3x 2. This problem can be described as the following math model: 1.2 Load Balancing Problem max z = 2x 1 + 3x 2 s.t. x 1 + 2x 2 8 4x x 2 12 x 1, x 2 0 For n processors with loaded work, distribute the new work such that the lightest-loaded processor has as heavy a load as possible. p i = current load of processor i = 1, 2, n, L = additional total load to be distributed, x i = fraction of additional load L distributed to processor i, with x i 0 and n i=1 x i = 1, 1

2 τ = minimum of final loads after distribution of workload L. We can formulate this problem as follows: Where e = {1, 1,, 1} T max x,τ s.t. τ τe p + xl e T x = 1 x Resource Allocation Produce m types of products, by using n resources. Each unit of product i yields c i dollars in revenue, whereas each unit of resource j costs d j dollars. One unit of product i requires A ij units of resource j to manufacture, and a maximum of b j units of resource j are available How to allocate resources to product products to maximize the profit? y i = the number of unites of product i x j = the number of units of resource j consumed We have max x,y s.t. z = c T y d T x x = A T y x b x, y 0 noindent Where the jth equation of x = A T y is x j = A 1j y 1 + A 2j y A mj y m 1.4 Approximation& Fittting In Economy, Finance, Marketing and other fields, we need to analyze the factors influecing some metrics, or make some predictions. For example, GDP prediction in Figure 1. These are approximation and fitting problems. Figure 1: GDP Prediction 2

3 Given m data points (a i, b i ), i = 1,, m, where a i R n and b i R, build a model that predicts the value of b from the vector a. A linear model b = a T x is a popular one. We should choose a model that explains the avaiable data as best as possible, i.e. a model that results in small residuals (Figure 2). One possible way is to mimimize the largest residual, that is to minimize max bi a T i x i Figure 2: Approximation& Fitting The following is an equivalent linear programming formulation: here, the decision variables being z and x. An alternative formulation is to adopt the cost criterion The corresponding formulation is min z s.t. b i a T i x z i = 1,, m, b i + a T i x z i = 1,, m, m b i a T i x i=1 min z z m s.t. b i a T i x z i i = 1,, m, b i + a T i x z i i = 1,, m, In practice, the quadratic cost criterion m (b i a T i x)2 is often adopted ( least square fit ), which will be discussed in the later chapters. 1.5 Pattern Classification i=1 Credit risk management is a typical Pattern Classification problem. A sample data of credit card application: 3

4 Figure 3: Credit Card Application In classification problems, given two sets of points in the space of n dimensions R n, find a hyperplane that separate these two set as accurately as possible. Let s see how to use linear programming to find the separating hyperplane. The hyperplane is defined by a vector ω R n and a scalar τ. Ideally, each point t in the first set satisfies ω T t τ, and one in the second set satisfies ω T t τ. To guard against a trivial answer (e.g. ω = 0 and τ 0, the conditions are trivially satisfied), we enforce the stronger conditions ω T t τ + 1 for points in the first set, and ω T t τ 1 for points in the second set. Figure 4: Pattern Classification Let M to be m n matrix, whose ith row contains the n components of the ith points in the first set. Similarly construct k n matrix B from the points in the second set. The violations of the condition ω T t τ + 1 for the point in the first set are measured by a vector y, which is defined by y (Mω τe) + e (here y 0 and e = (1, 1,, 1) T R m ). Similarly, for the points in second set, the violations are measured by z defined by z (Bω τe) + e (z 0, e R k ). The average violation on the first set is e T y/m and on the second set is e T z/k. The classification is formulated as follows: 4

5 min ω,τ,y,z s.t. e T y/m + e T z/m y (Mω τe) + e z (Bω τe) + e (y, z) 0 Figure 5: Classifier with Linear Programming 2 Linear Programming (LP) 2.1 General Form Given a cost vector c = (c 1,, c n ) T, we seek to minimize (maximize) a linear cost function c T x = n c i x i over all n-dimensional vector x = (x 1,, x n ) T, subject to a set of linear equality and inequality constraints. min c 1 x 1 + c 2 x c n x n s.t. a 11 x 1 + a 12 x a 1n x n b 1 a 21 x 1 + a 22 x a 2n x n b 2 a m11x 1 + a m12x a m1nx n b m1 i=1 a 11 x 1 + a 12 x a 1n x n b 1 a 21 x 1 + a 22 x a 2n x n b 2 a m21x 1 + a m22x a m2nx n b m 2 a 11 x 1 + a 12 x a 1n x n = b 1 a 21 x 1 + a 22 x a 2n x n = b 2 a m31x 1 + a m32x a m3nx n = b m 3 x 1, x 2,, x n ( )0 5

6 min c T x s.t. a T i x b i i M 1 a T i x b i i M 2 a T i x = b i i M 3 x j 0 j N 1 x j 0 j N 2 The variable x 1,, x n are called decision variables, a vector x satisfying all of the constraints is called a feasible solution. c t x is called the objective function, a feasible solusion x that minimizes the objective function (that is, c T x c T x for all feasible x) is called an optimal solution. If we can find a feasible solution x whose cost is less than any real number, we say that the optimal cost is (unbounded below, the problem is unbounded). 2.2 Reduction to standard form The standard form of a linear programming(lp) problem is min c 1 x 1 + c 2 x c n x n s.t. a 11 x 1 + a 12 x a 1n x n = b 1 a 21 x 1 + a 22 x a 2n x n = b 2 a m1 x 1 + a m2 x a mn x n = b m x 1, x 2,, x n 0 A general linear programming problem can be transformed into an equivalent problem in standard form: a) Elimination of free variables: Given an unrestricted variable x j, replace it by x + j x j, where x+ j 0, x j 0 b) Elimination of inequality constraints: Given an inequality constraint of the form n a ij x j ( )b i, introduce a new variable s i and convert the constrain as j=1 here, s i is called as a slack variable n a ij x j + s i ( s i ) = b i ; s i 0 j=1 Example 1. The linear programming problem can be converted into the stardard form min 2x 1 + 4x 2 s.t. x 1 + x 2 3 3x 1 + 2x 2 = 14 x 1 0 min 2x 1 + 4x + 2 4x 2 s.t. x 1 + x + 2 x 2 x 3 = 3 3x 1 + 2x + 2 2x 2 = 14 x 1, x + 2, x 2, x 3 0 6

7 3 References 1. Dimitris Bertsimas, John N. Tsitsiklis. Introduction to Linear Programming, Athena Scientific, Michael C. Ferris, Olvi L. Mangasarian, Stephen J. Wright. Linear Programming with Matlab, SIAM,