Presentation is loading. Please wait.

Presentation is loading. Please wait.

Hardness of Shops and Optimality of List Scheduling

Similar presentations


Presentation on theme: "Hardness of Shops and Optimality of List Scheduling"— Presentation transcript:

1 Hardness of Shops and Optimality of List Scheduling
Ola Svensson KTH Royal Institute of Technology Stockholm, Sweden

2 Gap Instance with r=3 and F=3

3 Jobs of different frequency “cannot” overlap
Key-Lemma: Jobs of different frequencies overlap at most a fraction 1/r of their length.

4 Reduction 1 2 Jobs corresponding to an IS can be scheduled in parallel since no short-jobs interfer with the long-jobs Jobs can be scheduled in parallel iff they correspond to an IS Adjacent jobs cannot overlap by Key-Lemma

5 Completeness: If Graph is K-colorable then there is a schedule of makespan at most K*lb Soundness: If Graph has no IS of size n/Klog(K) then any schedule has makespan at least Klog(K) *lb Some point

6 Bounding #frequencies
Take “enough” and assign random frequencies to jobs For a fixed K the graphs have bounded degree (Color in polytime using degree+1 colors, each color only needs one frequency)

7 Final Result SAT Graph Coloring/IS Job Shops makespan K lb K colors no IS of size n/KlogK makespan klogK lb It is NP-hard to approximate job shops within any constant factor Assuming NP-complete problems have no randomized quasi-polynomial then job shops have no better than log(lb) approximation.

8 Performance Guarantee
Gap Instance Hardness Result Job Shops Õ(log2(lb)) Shmoys, Stein, Wein’94 Goldberg,Paterson,Srinivisan, Sweedyk’97 Ω(log(lb)) Feige & Scheideler’98 Mastrolilli,S’08 Acyclic Job Shops Õ(log(lb)) Generalized Flow Shops Mastrolilli,S’09 Flow Shops Mastrolilli, S’09 5/4 Williamson, Hall, Hoogeven, Lenstra, Sevastianov, Shmoys’97 ~ ~ ~ ~ ~ ~ ~

9 Final Comments PTAS iff #machines and #operations per job bounded by a constant Jansen, Solis-Oba & Sviridenko’99 + Mastrolilli, S’08 Preemptive case wide open No non-constant gap instances Best hardness 5/4 and no constant approx algos.

10 Expander Not Expander Not scheduled 0.99n machines 0.99n machines

11 Not Expander Expander 2d d+1 0.99n machines 0.99n machines

12 Building block (an instance of 1|prec|ΣwjCj) Instance of P|prec| Cmax
processing time = 0 processing time = 1

13 Yes Case No Case 2d d+1 0.99n machines 0.99n machines

14 Do not have Two extreme cases:
almost all predecessors have been completed almost none of them have been completed

15 Final Result SAT Node Expansion P|prec|Cmax makespan d+1 makespan 2d Assuming 1|prec|ΣwjCj is hard to approximate within a factor less than 2 Then P|prec|Cmax is hard to approximate within a factor less than 2

16 Is 1|prec|ΣwjCj difficult? (1/2)
Special Case of Vertex Cover Removing Fixed Cost then equivalent to Vertex Cover No PTAS unless NP-complete problems can be solved in randomized polynomial time Correa & Schulz’04 , Ambuhl & Mastrolilli’06

17 Is 1|prec|ΣwjCj difficult? (2/2)
Assuming a variant of the Unique Games Conjecture It is NP-hard to approximate better than 2 Bansal & Khot’09 SAT Unique Games P|prec|Cmax makespan d+1 makespan 2d 1|prec|ΣwjCj Vertex Cover, Max Cut, Feedback arc set, rich family of CSPs…

18 Summary Non-constant hardness for job shops that are essentially tight for acyclic job shops and general flow shops Grahams list scheduling algorithm is optimal assuming a variant of the unique games conjecture Master plan has been formalized for some classes of problems Integrality gap of an SDP for CSPs implies UG-hardness Integrality gap of an LP for packing and covering problems implies UG-hardness Raghavendra’08 Kumar, Manokaran, Tulsiani, Vishnoi’11

19 Two open problems Scheduling precedence-constrained jobs on related machines Q|prec|Cmax log(m) approximation No better hardness than for parallel machines Scheduling on unrelated machines R||Cmax 2 approximation NP-hard to do better than 1.5 Chudak & Shmoys’97 Lenstra, Shmoys & Tardos’90

20


Download ppt "Hardness of Shops and Optimality of List Scheduling"

Similar presentations


Ads by Google