Outline. Relaxation. Outline DMP204 SCHEDULING, TIMETABLING AND ROUTING. 1. Lagrangian Relaxation. Lecture 12 Single Machine Models, Column Generation

Transcription

1 Outline DMP204 SCHEDULING, TIMETABLING AND ROUTING 1. Lagrangian Relaxation Lecture 12 Single Machine Models, Column Generation 2. Dantzig-Wolfe Decomposition Dantzig-Wolfe Decomposition Delayed Column Generation Marco Chiarandini Slides from David Pisinger s lectures at DIKU 3. Single Machine Models 2 Outline Relaxation In branch and bound we find upper bounds by relaxing the problem 1. Lagrangian Relaxation Relaxation { max g(s) maxs P f(s) s P max s S g(s) } max s S f(s) 2. Dantzig-Wolfe Decomposition Dantzig-Wolfe Decomposition Delayed Column Generation 3. Single Machine Models P : candidate solutions; S P feasible solutions; g(x) f(x) Which constraints should be relaxed? Quality of bound (tightness of relaxation) Remaining problem can be solved efficiently Proper multipliers can be found efficiently Constraints difficult to formulate mathematically Constraints which are too expensive to write up 3 4

2 Tightness of relaxation Different relaxations LP-relaxation Tighter max cx s.t. Ax b Dx d Deleting constraint Lagrange relaxation Surrogate relaxation Semidefinite relaxation Relaxations are often used in combination. Best surrogate relaxation Best Lagrangian relaxation LP relaxation x Z n + LP-relaxation: max {cx : x conv(ax b, Dx d, x Z + )} Lagrangian Relaxation: max z LR (λ) = cx λ(dx d) s.t. Ax b x Z n + LP-relaxation: max {cx : Dx d, x conv(ax b, x Z + )} 5 6 Relaxation strategies Which constraints should be relaxed "the complicating ones" remaining problem is polynomially solvable (e.g. min spanning tree, assignment problem, linear programming) remaining problem is totally unimodular (e.g. network problems) remaining problem is NP-hard but good techniques exist (e.g. knapsack) constraints which cannot be expressed in MIP terms (e.g. cutting) constraints which are too extensive to express (e.g. subtour elimination in TSP) 7 Subgradient optimization Lagrange multipliers max z = cx s. t. Ax b Dx d x Z n + Lagrange Relaxation, multipliers λ 0 max z LR (λ) = cx λ(dx d) s. t. Ax b x Z n + Lagrange Dual Problem z LD = min λ 0 z LR(λ) We do not need best multipliers in B&B algorithm Subgradient optimization fast method Works well due to convexity Roots in nonlinear programming, Held and Karp (1971) 8

3 Subgradient optimization, motivation Subgradient Generalization of gradients to non-differentiable functions. Definition An m-vector γ is subgradient of f(λ) at λ λ if f(λ) f( λ) + γ(λ λ) The inequality says that the hyperplane y = f( λ) + γ(λ λ) is tangent to y = f(λ) at λ λ and supports f(λ) from below Netwon-like method to minimize a function in one variable Lagrange function z LR (λ) is piecewise linear and convex 9 10 Proposition Given a choice of nonnegative multipliers λ. If x is an optimal solution to z LR (λ) then is a subgradient of z LR (λ) at λ = λ. γ = d Dx Proof We wish to prove that from the subgradient definition: max (cx = λ(dx d)) γ(λ λ) ( ) + max cx λ(dx d) Ax b Ax b where x is an opt. solution to the right-most subproblem. Inserting γ we get: max (cx λ(dx d)) (d Ax b Dx )(λ λ) + (cx λ(dx d)) = cx λ(dx d) Intuition Lagrange relaxation max z LR (λ) = cx λ(dx d) s.t. Ax b x Z n + Gradient in x is Subgradient Iteration Recursion where θ > 0 is step-size γ = d Dx λ k+1 = max { λ k θγ k, 0 } If γ > 0 and θ is sufficiently small z LR (λ) will decrease. Small θ slow convergence Large θ unstable 11 12

4 Lagrange relaxation and LP For an LP-problem where we Lagrange relax all constraints Dual variables are best choice of Lagrange multipliers Lagrange relaxation and LP "relaxation" give same bound Gives a clue to solve LP-problems without Simplex Iterative algorithms Polynomial algorithms Outline Dantzig-Wolfe Decomposition Motivation split it up into smaller pieces a large or difficult problem 1. Lagrangian Relaxation 2. Dantzig-Wolfe Decomposition Dantzig-Wolfe Decomposition Delayed Column Generation 3. Single Machine Models Applications Cutting Stock problems Multicommodity Flow problems Facility Location problems Capacitated Multi-item Lot-sizing problem Air-crew and Manpower Scheduling Vehicle Routing Problems Scheduling (current research) Two currently most promising directions for MIP: Branch-and-price Branch-and-cut 15 17

5 Dantzig-Wolfe Decomposition The problem is split into a master problem and a subproblem + Tighter bounds + Better control of subproblem Model may become (very) large Delayed column generation Write up the decomposed model gradually as needed Generate a few solutions to the subproblems Solve the master problem to LP-optimality Use the dual information to find most promising solutions to the subproblem Extend the master problem with the new subproblem solutions. 18

6

7 Delayed Column Generation Delayed column generation, linear master Master problem can (and will) contain many columns To find bound, solve LP-relaxation of master Delayed column generation gradually writes up master 29

8 Reduced Costs Simplex in matrix form min {cx Ax = b, x } In matrix form: [ ] [ ] 0 A z 1 c x B = {1, 2,..., p} basic variables = [ ] b 0 L = {1, 2,..., q} non-basis variables (will be set to lower bound = 0) (B, L) basis structure x B, x L, c B, c L, B = [A 1, A 2,..., A p ], L = [A p+1, A p+2,..., A p+q ] [ ] z [ 0 B L x 1 c B c B b = L 0] x L Bx B + Lx L = b x B + B 1 Lx L = B 1 b [ xl = 0 x B = B 1 b 31 [ ] z 0 B L x 1 c B c B = L x L Simplex algorithm sets x L = 0 and x B = B 1 b B invertible, hence rows linearly independent [ ] b 0 The objective function is obtained by multiplying and subtracting constraints by means of multipliers π (the dual variables) [ ] [ ] p p q p p z = c j π i a ij + c j π i a ij + π i b i i=1 i=1 Each basic variable has cost null in the objective function p c j π i a ij = 0 = π = B 1 c B i=1 Reduced costs of non-basic variables: p c j π i a ij i=1 i=1 32

9 Questions Will the process terminate? Always improving objective value. Only a finite number of basis solutions. Can we repeat the same pattern? No, since the objective functions is improved. We know the best solution among existing columns. If we generate an already existing column, then we will not improve the objective. 36 Outline Scheduling 1. Lagrangian Relaxation Sequencing (linear ordering) variables 1 prec P w j C j 2. Dantzig-Wolfe Decomposition Dantzig-Wolfe Decomposition Delayed Column Generation 3. Single Machine Models n n n min w j p k x kj + w j p j k=1 s.t. x kj + x lk + x jl 1 j, k, l = 1,..., nj k, k l x kj + x jk = 1 j, k = 1,..., n, j k x jk {0, 1} j, k = 1,..., n x jj = 0 j = 1,..., n 38 39

10 Scheduling Scheduling Time indexed variables 1 P h j (C j ) Completion time variables n min w j z j s.t. z k z j p k for j k A z j p j, for j = 1,..., n z k z j p k or z j z k p j, for (i, j) I z j R, j = 1,..., n 1 prec C max min s.t. n n h j (t + p j )x jt x jt = 1, t s=t p j +1 x js 1, for all j = 1,..., n for each t = 1,..., T x jt {0, 1}, for each j = 1,..., n; t = 1,..., T p j This formulation gives better bounds than the two preceding pseudo-polynomial number of variables Dantzig-Wolfe decomposition Reformulation: Dantzig-Wolfe decomposition Substituting X in original model getting master problem min s.t. n h j (t + p j )x jt x jt = 1, for all j = 1,..., n x jt X for each j = 1,..., n; t = 1,..., T p j + 1 n where X = x {0, 1} : t s=t p j +1 x js 1, for each t = 1,..., T π α min s.t. n h j (t + p j )( L λ l x l ) l=1 L λ l x l jt = 1, for all j = 1,..., n = l=1 L λ l = 1, l=1 λ l {0, 1} = λ l 0 LP-relaxation L λ l n l j = 1 l=1 x l, l = 1,..., L extreme points of X. x {0, 1} : x = L l=1 λ lx l X = L l=1 λ l = 1, λ l {0, 1} matrix of X is interval matrix extreme points are integral they are pseudo-schedules 42 solve LP-relaxation by column generation on pseudo-schedules x l reduced cost of λ k is c k = n (c jt π j )x k jt α 43

11 The subproblem can be solved by finding shortest path in a network N with 1, 2,..., T + 1 nodes corresponding to time periods process arcs, for all j, t, t t + p j and cost c jt π j idle time arcs, for all t, t t + 1 and cost 0 a path in this network corrsponds to a pseudo-schedule in which a job may be started more than once or not processed. the lower bound on the master problem produced by the LP-relaxation of the restricted master problem can be tighten by inequalities [Pessoa, Uchoa, Poggi de Aragão, Rodrigues, 2008], propose another time index formulation that dominates this one. They can solve consistently instances up to 100 jobs. 44