Chapter 6. Large Scale Optimization

Slides:



Advertisements
Similar presentations
Column Generation n ENGG 6070 n Instructor: Shawki M Areibi n Students: Limin Ma, Hao Qin.
Advertisements

Sandia is a multiprogram laboratory operated by Sandia Corporation, a Lockheed Martin Company, for the United States Department of Energy under contract.
Daniel Kroening and Ofer Strichman Decision Procedures An Algorithmic Point of View Deciding ILPs with Branch & Bound ILP References: ‘Integer Programming’
Decision Procedures An Algorithmic Point of View
1.3 Modeling with exponentially many constr.  Some strong formulations (or even formulation itself) may involve exponentially many constraints (cutting.
Chapter 3. Pitfalls Initialization Ambiguity in an iteration
Linear Programming Chapter 6. Large Scale Optimization.
3.3 Implementation (1) naive implementation (2) revised simplex method
OR Chapter 1. Introduction  Ex : Diet Problem Daily requirements : energy(2000kcal), protein(55g), calcium(800mg) Food Serving size Energy (kcal)
Column Generation By Soumitra Pal Under the guidance of Prof. A. G. Ranade.
Gomory Cuts Updated 25 March Example ILP Example taken from “Operations Research: An Introduction” by Hamdy A. Taha (8 th Edition)“Operations Research:
Branch-and-Cut Valid inequality: an inequality satisfied by all feasible solutions Cut: a valid inequality that is not part of the current formulation.
Implicit Hitting Set Problems Richard M. Karp Erick Moreno Centeno DIMACS 20 th Anniversary.
OR Chapter 8. General LP Problems Converting other forms to general LP problem : min c’x  - max (-c)’x   = by adding a nonnegative slack variable.
1.2 Guidelines for strong formulations  Running time for LP usually depends on m and n ( number of iterations O(m), O(log n)). Not critically depend on.
Chapter 2. Optimal Trees and Paths Combinatorial Optimization
Chapter 10 Advanced Topics in Linear Programming
OR Chapter 4. How fast is the simplex method  Efficiency of an algorithm : measured by running time (number of unit operations) with respect to.
Linear Programming Piyush Kumar Welcome to CIS5930.
Integer Programming An integer linear program (ILP) is defined exactly as a linear program except that values of variables in a feasible solution have.
Chap 10. Sensitivity Analysis
Linear Programming Topics General optimization model
Chapter 1. Introduction Ex : Diet Problem
Perturbation method, lexicographic method
6.5 Stochastic Prog. and Benders’ decomposition
EMGT 6412/MATH 6665 Mathematical Programming Spring 2016
Proving that a Valid Inequality is Facet-defining
The Two-Phase Simplex Method
Linear Programming Topics General optimization model
Gomory Cuts Updated 25 March 2009.
1.3 Modeling with exponentially many constr.
Linear Programming Topics General optimization model
Chap 9. General LP problems: Duality and Infeasibility
Polyhedron Here, we derive a representation of polyhedron and see the properties of the generators. We also see how to identify the generators. The results.
Chapter 6. Large Scale Optimization
Chapter 5. Sensitivity Analysis
Chap 3. The simplex method
Integer Programming (정수계획법)
Polyhedron Here, we derive a representation of polyhedron and see the properties of the generators. We also see how to identify the generators. The results.
Chapter 4. Duality Theory
2. Generating All Valid Inequalities
Cutting Stock Problem Problem of cutting an unlimited number of pieces of material (paper rolls, for instance) of length l to produce ni pieces of length.
Chapter 8. General LP Problems
2.2 Shortest Paths Def: directed graph or digraph
1.3 Modeling with exponentially many constr.
Chapter 5. The Duality Theorem
Integer Programming (정수계획법)
I.4 Polyhedral Theory (NW)
Flow Feasibility Problems
Cynthia Phillips (Sandia National Laboratories)
Back to Cone Motivation: From the proof of Affine Minkowski, we can see that if we know generators of a polyhedral cone, they can be used to describe.
I.4 Polyhedral Theory.
Proving that a Valid Inequality is Facet-defining
Chapter 8. General LP Problems
(Convex) Cones Def: closed under nonnegative linear combinations, i.e.
Chapter-III Duality in LPP
6.5 Stochastic Prog. and Benders’ decomposition
Chapter 2. Simplex method
Simplex method (algebraic interpretation)
DUALITY THEORY Reference: Chapter 6 in Bazaraa, Jarvis and Sherali.
BASIC FEASIBLE SOLUTIONS
Chapter 8. General LP Problems
Branch-and-Bound Algorithm for Integer Program
1.2 Guidelines for strong formulations
Chapter 6. Large Scale Optimization
Chapter 2. Simplex method
1.2 Guidelines for strong formulations
Chapter 3. Pitfalls Initialization Ambiguity in an iteration
Discrete Optimization
Presentation transcript:

Chapter 6. Large Scale Optimization 6.1 Delayed column generation min c’x, Ax = b, x  0. A is full row rank with a large number of columns. Impractical to have all columns initially. Want to generate (find) entering nonbasic variable (column) as needed. ( column generation technique ) If ci < 0, then xi can enter basis. Hence solve min ci over all i. If min ci < 0, have found an entering variable (column). If min ci  0, no entering column exists, hence current basis optimal. If entering col. found, solve restricted problem to optimality. min i  I , i  I Aixi = b, x  0 ( I : index set of variables we have at hand) Then continue to find entering columns. Linear Programming 2011

6.2. Cutting stock problem W = 70 17 17 17 15 scrap Linear Programming 2011

bi rolls of width wi , i = 1, 2, … , m need to be produced. Rolls of paper with width W ( called raw) to be cut into small pieces (called final). bi rolls of width wi , i = 1, 2, … , m need to be produced. How to cut the raws to minimize the number of raws used while satisfying order? ex) W = 70, then 3 of w1 = 17 and 1 of w2 = 15 can be produced from a raw. This way of production can be represented as pattern ( 3, 1, 0, 0, … , 0 ) ( a1j, a2j, … , amj )’ for j-th pattern is feasible if i aijwi  W. Linear Programming 2011

, where aij is the number of i-th finals produced in j-th pattern. Formulation , where aij is the number of i-th finals produced in j-th pattern. Note that the number of possible cutting patterns can be very large. We need integer solution, but LP relaxation can be used to find good approximate solution if solution value large. For initial b.f.s., for j = 1, … , m, let j-th pattern consists of one final of width wj and none of the other widths. Linear Programming 2011

 max p’Aj over all patterns ( integer knapsack problem ) After computing p vector, we try to find entering nonbasic variable (column). Candidate for entering column (pattern) is any nonbasic variable with reduced cost ( 1 – p’Aj ) < 0, hence solve min ( 1 – p’Aj ) over all possible patterns.  max p’Aj over all patterns ( integer knapsack problem ) max Dynamic programming algorithm for integer knapsack prob. ( can assume pi, wi > 0, integer. knapsack is NP-hard, so no polynomial time alg. is known. ) Linear Programming 2011

Let F(v) be optimal value of the problem when knapsack capacity is v. wmin = mini { wi } For v < wmin , F(v) = 0 For v  wmin , F(v) = maxi = 1, … , m { F( v – wi ) + pi : v  wi } F(v) is true optimal value when knapsack capacity is v. Suppose a0 is opt. solution when r.h.s. is v – wi , then a0 + ei is a feasible solution when r.h.s. is v. Hence F(v)  F(v – wi) + pi , i = 1, … , m, v  wi Suppose a* is optimal solution when r.h.s. is v  wmin , then there exists some k with a*k > 0 and v  wk . Hence a* - ek is a feasible solution when r.h.s. is v – wk . So F( v-wk )  F(v) – pk ( F(v)  F( v-wk ) + pk for some k ) Linear Programming 2011

Actual solution recovered by backtracking the recursion. Running time of the algorithm is O(mW) which is not polynomial of the length of encoding. Called pseudopolynomial running time ( polynomial of data W itself). Note : the running time becomes polynomial if it is polynomial with respect to m and log2 W, but W = 2logW which is not polynomial of log2 W. Many practical problems can be naturally formulated similar to the cutting stock problem. Especially in 0-1 IP with many columns. For cutting stock problem, we only obtained approximate fractional solution. But for 0-1 IP, fractional solution can be of little help and we need a mechanism to find optimal integer solution ( branch-and-price approach, column generation combined with branch-and-bound ). Linear Programming 2011

6.3. Cutting plane methods Dual of column generation (constraint generation) Consider max p’b, p’Ai  ci , i = 1, … , n (1) ( n can be very large ) Solve max p’b, p’Ai  ci , i  I, I  { 1, … , n } (2) and get optimal solution p* to (2) If p* is feasible to (1), then it is also optimal to (1) If p* is infeasible to (1), find a violated constraint in (1) and add it to (2), then reoptimize (2) again. Repeat it. Recall the TSP formulation with subtour elimination constraints. Linear Programming 2011

Solve min ci – (p*)’Ai over all i. If optimal value  0  p*  P Separation problem : Given a polyhedron P (described with possibly many inequalities) and a vector p* , determine if p*  P. If p*  P, find a (valid) inequality violated by p*. Solve min ci – (p*)’Ai over all i. If optimal value  0  p*  P If optimal value < 0  ci < (p*)’Ai (violated) Linear Programming 2011

6.4. Dantzig-Wolfe decomposition Use of decomposition theorem to represent a specially structured LP problem in different form. Column generation is used to solve the problem. Consider a LP in the following form min x1, x2 : dimension n1, n2, b0, b1, b2 : m0, m1, m2 Let Pi = { xi  0 : Fix = bi }, i = 1, 2 Assume Pi  . Note that the nonnegativity constraints guarantee that Pi is pointed, hence S = {0} ( P = S + K + Q ) Linear Programming 2011

xi  Pi can be represented as  min xi  Pi can be represented as Plug into (2)  get master problem min Linear Programming 2011

Alternatively, its columns can be viewed as The new formulation has many variables (columns), but it can be solved by column generation technique. Actual solution x1, x2 can be recovered from  and . xi is expressed as convex combination of extreme points of Pi + conical combination of extreme rays of Pi. Linear Programming 2011

Decomposition algorithm Suppose having a b.f.s. to the master problem, dual value p = ( q, r1, r2), qRm0, r1, r2  R. Then reduced costs are (for 1, 1 ) Entering variable if reduced cost < 0. Hence solve min ( c1’ – q’D1 )x1, x1  P1 (subproblem) Linear Programming 2011

simplex returns extreme ray w1k with ( c1’ – q’D1 )w1k < 0. (a) optimal cost is -   simplex returns extreme ray w1k with ( c1’ – q’D1 )w1k < 0. Generate column for 1k, i.e. [ D1w1k ’, 0 , 0 ]’ (b) optimal finite and < r1  returns extreme point x1j with ( c1’ – q’D1 )x1j < r1. Generate column for 1j, i.e. [ D1x1j ’, 1 , 0 ]’ (c) optimal cost  r1  ( c1’ – q’D1 )x1j  r1  x1j , ( c1’ – q’D1 )w1k  0  w1k no entering variable among 1j , 1k Perform the same for 2j , 2k The method can also be used when there are more than 2 blocks or just one block in the constraints. Linear Programming 2011

Starting the algorithm Find extreme points x11, x21 of P1 and P2 May assume that D1x11 + D2x21  b, then solve. min Linear Programming 2011

Termination and computational experience Fast improvement in early iterations, but convergence becomes slow in the tail of the sequence. Revised simplex is more competitive in terms of running time. Suitable for large, structured problems. Linear Programming 2011

Bounds on the optimal cost Thm 6.1 : Suppose optimal z* is finite. Let z be the current best solution (upper bound on z* ), ri dual variable value for i-th convexity constraint and zi finite optimal cost for i-th subproblem. Then z + i ( zi – ri )  z*  z. pf) Modify the current dual solution to a dual feasible solution by decreasing the value of ri to zi . Dual of master problem is max Linear Programming 2011

Suppose have a b.f.s. to master problem with z and ( q, r1, r2 ). Have q’b0 + r1 + r2 = z Optimal cost z1 to the first subproblem finite  minj  J1 (c1’x1j – q’D1x1j ) = z1 mink  k1 (c1’w1k – q’D1w1k )  0 Note that currently we have minj ( c1’ – q’D1 )x1j < r1 ( reduced cost (c1’ – q’D1 )x1j - r1 < 0 for entering variable). If we use z1 in place of r1, get dual feasibility for the first two dual constraints. Similarly, use z2 in place of r2. Cost is q’b0 + z1 + z2  z*  q’b0 + z1 + z2 = q’b0 +r1 + r2 + ( z1 – r1 ) + ( z2 – r2 ) = z + ( z1 – r1 ) + ( z2 – r2 )  Linear Programming 2011