cumulative scheduling and pseudo boolean constraints
play

Cumulative Scheduling and Pseudo-Boolean Constraints Deduction at - PowerPoint PPT Presentation

Cumulative Scheduling and Pseudo-Boolean Constraints Deduction at Scale Albert Oliveras (+ Ignasi Ab o, Roberto As n, Robert Nieuwenhuis, Enric Rodr guez, Javier Larrosa, ...) Barcelogic Research Group, Tech. Univ. Catalonia,


  1. Cumulative Scheduling and Pseudo-Boolean Constraints Deduction at Scale Albert Oliveras (+ Ignasi Ab´ ıo, Roberto As´ ın, Robert Nieuwenhuis, Enric Rodr´ ıguez, Javier Larrosa, ...) Barcelogic Research Group, Tech. Univ. Catalonia, Barcelona Barcelogic - Tech. Univ. Catalonia (UPC) – p. 1

  2. Motivation Assume a company needs to complete the following 15 jobs: # Job Transport Unload Process Pack 1 1 truck 10 unloaders 5 technicians 3 packers 2 1 truck 5 unloaders 3 technicians, 1 supervisor 2 packers 3 1 car 2 unloaders 1 technicians , 1 supervisor 1 packer ... 15 1 truck 5 unloaders 3 technicians 2 packers Tasks cannot be interrupted (non-preemptive) Task precedences: transport → unload ❀ process ❀ pack Each task has a certain duration Resource availability: 2 trucks, 4 cars, 30 unloaders, 13 techn., 3 superv., 8 packers GOAL: complete all jobs as early as possible Barcelogic - Tech. Univ. Catalonia (UPC) – p. 2

  3. Overview of the talk Cumulative scheduling SAT encodings for PB constraints SAT encodings vs SMT Existing encodings Standard BDD-based encoding New results for BDD-based encodings CNFs for monotonic functions Probably BDDs for some PB constraints are non-poly Avoiding non-polynomial BDDs Conclusions and Future Work Barcelogic - Tech. Univ. Catalonia (UPC) – p. 3

  4. Problem statement INPUT DATA: For each resource r ∈ R , an integer avail ( r ) For each task t ∈ T : d ( t ) : duration of the task est ( t ) : earliest starting time lct ( t ) : latest completion time List of consumptions { ( r , units ( t , r )) , . . . , } List of precedences t 1 ❀ t 2 , and immediate ones t 1 → t 2 GOAL: find a valid schedule that minimizes makespan (i.e. minimizes the maximum completion time) Barcelogic - Tech. Univ. Catalonia (UPC) – p. 4

  5. Solving method Best approach: encoding to SAT + PB-constraints ([StuckeyCP09]). Consider propositional variables: s t , i ≡ start _ time ( t ) ≤ i and add clauses: s t ,0 → s t ,1 s t ,1 → s t ,2 . . . [integer consistency] ¬ s t , est ( t ) − 1 s t , lct ( t ) − d ( t ) [satisfy est and lct ] s t 2 ,0 → s t 1 ,0 s t 2 ,1 → s t 1 ,1 . . . [ t 1 precedes t 2 ] Barcelogic - Tech. Univ. Catalonia (UPC) – p. 5

  6. Solving method Best approach: encoding to SAT + PB-constraints ([StuckeyCP09]). Consider propositional variables: s t , i ≡ start _ time ( t ) ≤ i and add clauses: s t ,0 → s t ,1 s t ,1 → s t ,2 . . . [integer consistency] ¬ s t , est ( t ) − 1 s t , lct ( t ) − d ( t ) [satisfy est and lct ] s t 2 ,0 → s t 1 ,0 s t 2 ,1 → s t 1 ,1 . . . [ t 1 precedes t 2 ] Consider propositional variables: a t , i ≡ task t is active at time i and define them with: ↔ start _ time ( t ) ≤ i ∧ start _ time ( t ) + d ( t ) ≥ i + 1 a t , i Barcelogic - Tech. Univ. Catalonia (UPC) – p. 5

  7. Solving method Best approach: encoding to SAT + PB-constraints ([StuckeyCP09]). Consider propositional variables: s t , i ≡ start _ time ( t ) ≤ i and add clauses: s t ,0 → s t ,1 s t ,1 → s t ,2 . . . [integer consistency] ¬ s t , est ( t ) − 1 s t , lct ( t ) − d ( t ) [satisfy est and lct ] s t 2 ,0 → s t 1 ,0 s t 2 ,1 → s t 1 ,1 . . . [ t 1 precedes t 2 ] Consider propositional variables: a t , i ≡ task t is active at time i and define them with: ↔ start _ time ( t ) ≤ i ∧ start _ time ( t ) + d ( t ) ≥ i + 1 a t , i Add PB-constraints of the form: ∑ units ( t , r ) ∗ a t , i ≤ avail ( r ) , for each r and time i t s.t. t uses r Barcelogic - Tech. Univ. Catalonia (UPC) – p. 5

  8. Solving method Best approach: encoding to SAT + PB-constraints ([StuckeyCP09]). Consider propositional variables: s t , i ≡ start _ time ( t ) ≤ i and add clauses: s t ,0 → s t ,1 s t ,1 → s t ,2 . . . [integer consistency] ¬ s t , est ( t ) − 1 s t , lct ( t ) − d ( t ) [satisfy est and lct ] s t 2 ,0 → s t 1 ,0 s t 2 ,1 → s t 1 ,1 . . . [ t 1 precedes t 2 ] Consider propositional variables: a t , i ≡ task t is active at time i and define them with: ↔ start _ time ( t ) ≤ i ∧ start _ time ( t ) + d ( t ) ≥ i + 1 a t , i Add PB-constraints of the form: 2 x 1 + 7 x 2 + 5 x 3 + 10 x 4 ≤ 12 Barcelogic - Tech. Univ. Catalonia (UPC) – p. 5

  9. Overview of the talk Cumulative scheduling SATencodings for PB constraints SAT encodings vs SMT Existing encodings Standard BDD-based encoding New results for BDD-based encodings CNFs for monotonic functions Probably BDDs for some PB constraints are non-poly Avoiding non-polynomial BDDs Conclusions and Future Work Barcelogic - Tech. Univ. Catalonia (UPC) – p. 5

  10. SAT Encoding for PB-Constraints SAT + PB-constraints problems solvable with an SMT system But ... recent experience with SAT + cardinality constraints (PB with all coeffs. 1): Speed-up factor if encoding Slow-down factor if encoding Suite TO 4 2 1.5 TOT. 1.5 2 4 TO TOT. 168 54 14 7 243 7 24 215 12 258 MSU4 DES 321 53 20 16 410 11 4 0 0 15 MSU4: unsat-based max-SAT solving DES: discrete event system diagnosis problems How to decide whether SAT or SMT is the best choice? Why can SAT sometimes beat SMT? Barcelogic - Tech. Univ. Catalonia (UPC) – p. 6

  11. How could SAT beat SMT? SMT solvers can generate unsat proofs, which come in two parts: A resolution refutation from: the clauses of the input CNF the generated explanations (clauses) For each explanation clause, an independent proof in (its) T . So, after all, SMT generates a SAT encoding, but lazily. SMT solver runtime ≥ size of smallest resolution proof. Barcelogic - Tech. Univ. Catalonia (UPC) – p. 7

  12. How could SAT beat SMT? (2) In “artificial-like” problems: SMT’s lazy SAT encoding could end up being a full one And... this full encoding could be a rather naive one. Example: T = cardinality constraints. T -solver is just a counter. x 1 + . . . + x n ≤ k and x 1 + . . . + x n > k Unsat instance: Refutation requires all ( n k + 1 ) explanations like, e.g., x 1 ∨ x 2 ∨ . . . ∨ x k + 1 Here a good SAT encoding with auxiliary vars works better. Splitting on aux vars can give expon. speedup: Extended Resol. But... some constraints admit no P-size domain-consistent SAT encoding, e.g., alldiff [BessiereEtal’09]. Barcelogic - Tech. Univ. Catalonia (UPC) – p. 8

  13. PB-constraints SAT Encodings Desirable properties of a SAT Encoding: Polynomial size w.r.t. the number of vars of the constraint Consistent: take 2 x 1 + 4 x 2 + 2 x 3 ≤ 5 and I = { x 1 , x 2 } . Unit propagation should detect a false clause. Arc-consistent: take 2 x 1 + 3 x 2 + 7 x 3 ≤ 5 and I = { x 1 } . Unit propagation should propagate x 3 Existing encodings: Encoding Size Consistent Arc-consistent Tree-like Exponential yes yes BDD Not known yes yes O ( ∑ log a i ) Adders no no O (( ∑ log a i ) log 2 ∑ log a i ) Sorting yes no O ( n 2 log n log a max ) WatchDog yes yes Barcelogic - Tech. Univ. Catalonia (UPC) – p. 9

  14. Why yet another encoding? Existing arc-consistent polyn. encoding depends on coeffs. Take constraints: 3 x 1 + 2 x 2 + 4 x 3 ≤ 5 30001 x 1 + 19999 x 2 + 39998 x 3 ≤ 50007 They both represent the same Boolean function: ¬ ( x 1 ∧ x 3 ) ∧ ¬ ( x 2 ∧ x 3 ) 2nd constr. would have an unnecessarily large encoding A BDD-based encoding would solve this problem Question: what is the BDD size ? Barcelogic - Tech. Univ. Catalonia (UPC) – p. 10

  15. Overview of the talk Cumulative scheduling SAT encodings for PB constraints SAT encodings vs SMT Existing encodings Standard BDD-based encoding New results for BDD-based encodings CNFs for monotonic functions Probably BDDs for some PB constraints are non-poly Avoiding non-polynomial BDDs Conclusions and Future Work Barcelogic - Tech. Univ. Catalonia (UPC) – p. 10

  16. BDD construction process Take constraint 2 x 1 + 3 x 2 + 5 x 3 ≤ 6 a construct the diagram: x 1 0 1 x 2 x 2 0 1 0 1 x 3 x 3 x 3 x 3 0 1 0 1 0 1 0 1 1 0 Ordering x 1 < x 2 < x 3 is previously chosen Barcelogic - Tech. Univ. Catalonia (UPC) – p. 11

  17. BDD construction process Take constraint 2 x 1 + 3 x 2 + 5 x 3 ≤ 6 a construct the diagram: x 1 x 1 0 1 0 1 x 2 x 2 x 2 x 2 0 1 0 1 1 0 1 0 x 3 x 3 x 3 x 3 x 3 x 3 x 3 0 0 1 0 0 1 1 0 1 0 1 0 1 1 1 0 1 0 Nodes with identical children are removed Barcelogic - Tech. Univ. Catalonia (UPC) – p. 11

  18. BDD construction process Take constraint 2 x 1 + 3 x 2 + 5 x 3 ≤ 6 a construct the diagram: x 1 x 1 x 1 0 1 0 1 0 1 x 2 x 2 x 2 x 2 x 2 x 2 1 0 0 1 0 1 1 0 1 1 0 0 x 3 x 3 x 3 x 3 x 3 x 3 x 3 x 3 0 0 1 0 0 1 1 0 1 0 1 0 1 0 1 1 1 0 1 0 1 0 Isomorphic subtrees are merged Barcelogic - Tech. Univ. Catalonia (UPC) – p. 11

  19. BDD construction process Take constraint 2 x 1 + 3 x 2 + 5 x 3 ≤ 6 a construct the diagram: x 1 0 1 x 2 1 0 x 3 0 1 1 0 Canonical BDD for the constraint and ordering x 1 < x 2 < x 3 Barcelogic - Tech. Univ. Catalonia (UPC) – p. 11

  20. CNF conversion of BDDs x 1 0 1 x 2 1 0 x 3 0 1 1 0 Add auxiliary (output) variable for each node: o 1 , o 2 and o 3 Barcelogic - Tech. Univ. Catalonia (UPC) – p. 12

  21. CNF conversion of BDDs x 1 0 1 x 2 1 0 x 3 0 1 1 0 Add auxiliary (output) variable for each node: o 1 , o 2 and o 3 Specify behaviour of each node. For example: x 1 ∧ o 2 → o 1 x 1 ∧ o 3 → o 1 o 2 ∧ o 3 → o 1 x 1 ∧ o 2 → o 1 x 1 ∧ o 3 → o 1 o 2 ∧ o 3 → o 1 Barcelogic - Tech. Univ. Catalonia (UPC) – p. 12

Recommend


More recommend