Integrating Mixed-Integer Optimisation & Satisfiability Modulo Theories

Size: px

Start display at page:

Download "Integrating Mixed-Integer Optimisation & Satisfiability Modulo Theories"

Rosalind Stanley
5 years ago
Views:

1 Integrating Mixed-Integer Optimisation & Satisfiability Modulo Theories Application to Scheduling Miten Mistry and Ruth Misener Wednesday 11 th January, 2017 Mistry & Misener MIP & SMT Wednesday 11 th January, / 33

2 Optimisation & Logic Possible modelling frameworks: Generalised disjunctive programming, e.g. Grossmann & co-workers Mixed logical-linear programming, e.g. Hooker & co-workers Exisiting solution frameworks (Maravelias and Sung, 2009): Solve as a monolithic optimisation problem Solve as a monolithic satisfiability problem (Bjørner and De Moura, 2011) Decomposition methods e.g., Logic-based Benders decomposition (Jain and Grossmann, 2001; Hooker and Ottoson, 2003) Task 7 Task 4 Task 2 Machine 2 Task 6 Task 8 Task 1 Task 3 Task 5 Machine Time Mistry & Misener MIP & SMT Wednesday 11 th January, / 33

3 Background Mixed Integer Optimisation Mixed Integer Linear Programming (MILP) min c x s.t. Ax b x R n x j Z, j I {1,..., n} where c R n, A R m n and b R m. Mixed Integer Nonlinear Programming (MINLP) min f(x) s.t. g(x) 0 x R n x j Z, j I {1,..., n} where f : R n R, g : R n R m. Mistry & Misener MIP & SMT Wednesday 11 th January, / 33

4 Background Propositional Satisfiability (SAT) Alphabet soup: SAT CP SMT ASP Propositional Satisfiability Constraint Programming Satisfiability Modulo Theories Answer set programming Propositional Satisfiability (SAT) Analogy to machine code Given a propositional formula ϕ (built from, and ) over a set of propositional variables, p i, i I = {1,..., n}, can we find an assignment on the p i s such that ϕ evaluates to True (satisfied)? Other Constraint Satisfaction Techniques Analogy to high level language CP, SMT, ASP Mistry & Misener MIP & SMT Wednesday 11 th January, / 33

5 Background Constraint Programming (CP) Special constraints exploit structure found in specific applications Example: cumulative constraint mainly used in scheduling applications s: start times p: processing times c: resource consumptions C: resource consumption limit Hybridisation with MILP cumulative(s, p, c, C) Bockmayr and Kasper (1998); Jain and Grossmann (2001); Hooker and Ottoson (2003); Maravelias and Grossmann (2004); Bockmayr and Kasper (2004); Hooker (2007), etc. Mistry & Misener MIP & SMT Wednesday 11 th January, / 33

6 Background Satisfiability Modulo Theories (SMT) Reason about satisfiability (SAT) with respect to a theory Encode the SMT formula into a propositional formula and use the theory solver with the SAT solver to assess satisfiability. Can result in: Theories of Interest SAT UNSAT UNKNOWN Satisfiable Unsatisfiable For our applications usually a timeout Integer (Real) arithmetic - Interpret the symbols (0, 1, +,, =, ) in the usual way over Z (R) Arrays Combined theories - e.g. combining integer arithmetic with arrays to reason about a[i+j] Mistry & Misener MIP & SMT Wednesday 11 th January, / 33

7 Background Example: Satisfiability Modulo Theories (SMT) Example: Satisfiability Modulo Theories SAT MINLP meets SMT Callia D Iddio, Huth, Misener Mistry & Misener MIP & SMT Wednesday 11 th January, / 33

8 Background Example: Satisfiability Modulo Theories (SMT) Example: Satisfiability Modulo Theories UNSAT MINLP meets SMT Callia D Iddio, Huth, Misener Mistry & Misener MIP & SMT Wednesday 11 th January, / 33

9 Background Example: Satisfiability Modulo Theories (SMT) Example: Satisfiability Modulo Theories WHY UNSAT? MINLP meets SMT Callia D Iddio, Huth, Misener Mistry & Misener MIP & SMT Wednesday 11 th January, / 33

MINLP meets SMT Callia D Iddio, Huth, Misener Mistry

10 Background Example: Satisfiability Modulo Theories (SMT) Example: Satisfiability Modulo Theories SAT MINLP meets SMT Callia D Iddio, Huth, Misener Mistry & Misener MIP & SMT Wednesday 11 th January, / 33

11 Background Example: Satisfiability Modulo Theories (SMT) Example: Satisfiability Modulo Theories 0.15 x + y x y x + y x y 1.50 x, y 2 Z x 2 [ 1, 2] x 2 [ 2, 2] 0.10 x 2 + y 0.00 SAT MINLP meets SMT Callia D Iddio, Huth, Misener Mistry & Misener MIP & SMT Wednesday 11 th January, / 33

12 Background Example: Satisfiability Modulo Theories (SMT) Example: Satisfiability Modulo Theories (set-option :produce-unsat-cores true) (declare-const x Int) (declare-const y Int) (assert (! (>= (+ (* 0.15 x) (* 1 y)) 0.35) :named a1)) (assert (! (>= (+ (* x) (* -1 y)) -0.75) :named a2)) (assert (! (>= (+ (* x) (* 1 y)) 0.20) :named a3)) (assert (! (>= (+ (* 0.15 x) (* -1 y)) -1.50) :named a4)) (push) (check-sat)! SAT rise4fun.com/z3 a2 a4 a3 a1 MINLP meets SMT Callia D Iddio, Huth, Misener Mistry & Misener MIP & SMT Wednesday 11 th January, / 33

13 Background Example: Satisfiability Modulo Theories (SMT) Example: Satisfiability Modulo Theories (set-option :produce-unsat-cores true) (declare-const x Int) (declare-const y Int) (assert (! (>= (+ (* 0.15 x) (* 1 y)) 0.35) :named a1)) (assert (! (>= (+ (* x) (* -1 y)) -0.75) :named a2)) (assert (! (>= (+ (* x) (* 1 y)) 0.20) :named a3)) (assert (! (>= (+ (* 0.15 x) (* -1 y)) -1.50) :named a4)) (push) (check-sat)! SAT (assert (! (>= x -1) :named a5)) (assert (! (<= x 2) :named a6)) (check-sat)! UNSAT a2 rise4fun.com/z3 a5 a6 a4 a3 a1 MINLP meets SMT Callia D Iddio, Huth, Misener Mistry & Misener MIP & SMT Wednesday 11 th January, / 33

14 Background Example: Satisfiability Modulo Theories (SMT) Example: Satisfiability Modulo Theories (set-option :produce-unsat-cores true) (declare-const x Int) (declare-const y Int) (assert (! (>= (+ (* 0.15 x) (* 1 y)) 0.35) :named a1)) (assert (! (>= (+ (* x) (* -1 y)) -0.75) :named a2)) (assert (! (>= (+ (* x) (* 1 y)) 0.20) :named a3)) (assert (! (>= (+ (* 0.15 x) (* -1 y)) -1.50) :named a4)) (push) (check-sat)! SAT (assert (! (>= x -1) :named a5)) (assert (! (<= x (check-sat) (get-unsat-core) 2) :named a6))! UNSAT! (a1 a2 a6 a5) a2 rise4fun.com/z3 a5 a6 a4 a3 a1 MINLP meets SMT Callia D Iddio, Huth, Misener Mistry & Misener MIP & SMT Wednesday 11 th January, / 33

15 Background Example: Satisfiability Modulo Theories (SMT) Example: Satisfiability Modulo Theories (set-option :produce-unsat-cores true) (declare-const x Int) (declare-const y Int) (assert (! (>= (+ (* 0.15 x) (* 1 y)) 0.35) :named a1)) (assert (! (>= (+ (* x) (* -1 y)) -0.75) :named a2)) (assert (! (>= (+ (* x) (* 1 y)) 0.20) :named a3)) (assert (! (>= (+ (* 0.15 x) (* -1 y)) -1.50) :named a4)) (push) (check-sat)! SAT (assert (! (>= x -1) :named a5)) (assert (! (<= x 2) :named a6)) (check-sat) (get-unsat-core) (pop) (assert (! (>= x -2) :named a7)) (assert (! (<= x (check-sat)! UNSAT! (a1 a2 a6 a5) 2) :named a8))! SAT a3 a2 rise4fun.com/z3 a1 a7 a8 a4 MINLP meets SMT Callia D Iddio, Huth, Misener Mistry & Misener MIP & SMT Wednesday 11 th January, / 33

16 Background Example: Satisfiability Modulo Theories (SMT) Example: Satisfiability Modulo Theories rise4fun.com/z3! SAT! UNSAT! (a1 a2 a6 a5) a2 a7 a8 a4 a1! SAT a3! SAT MINLP meets SMT Callia D Iddio, Huth, Misener Mistry & Misener MIP & SMT Wednesday 11 th January, / 33

17 Two Dimensional Bin Packing Two Dimensional Bin Packing (2BP) Mistry & Misener MIP & SMT Wednesday 11 th January, / 33

18 Two Dimensional Bin Packing Two Dimensional Bin Packing (2BP) Mistry & Misener MIP & SMT Wednesday 11 th January, / 33

19 Two Dimensional Bin Packing Two Dimensional Bin Packing (2BP) Mistry & Misener MIP & SMT Wednesday 11 th January, / 33

20 Two Dimensional Bin Packing Two Dimensional Bin Packing (2BP) Mistry & Misener MIP & SMT Wednesday 11 th January, / 33

21 Two Dimensional Bin Packing Two Dimensional Bin Packing (2BP) Mistry & Misener MIP & SMT Wednesday 11 th January, / 33

22 Two Dimensional Bin Packing Two Dimensional Bin Packing (2BP) Mistry & Misener MIP & SMT Wednesday 11 th January, / 33

23 Two Dimensional Bin Packing Two Dimensional Bin Packing (2BP) Mistry & Misener MIP & SMT Wednesday 11 th January, / 33

24 Two Dimensional Bin Packing Logical Formulation Given items j {1,..., n} with widths W j and heights H j. Given n bins with width W and height H. min s.t. n b=1 c b n ζ jb b=1 (ζ ib ζ jb ) b b ζ jb { (x i + W i x j ) (x j + W j x i ) ζ jb ζ b ζ b (c b = 1), ζ b (c b = 0) 0 x j W j, H j y j H (y i H i y j ) (y j H j y i ) Mistry & Misener MIP & SMT Wednesday 11 th January, / 33

25 Two Dimensional Bin Packing Mixed Integer Formulation min s.t. n b=1 z b n z jb = 1 b=1 x i + W i x j + M 1 ij(1 z 1 ij), x j + W j x i + M 2 ij(1 z 2 ij) y i H i y j Mij(1 3 zij), 3 y j H j y i Mij(1 4 zij) 4 4 n zij k = z ib z jb k=1 b=1 z b z jb 0 x j W j, H j y j H Big-M Formulation, No Symmetry Breaking Lodi et al. (2002); Trespalacios and Grossmann (2016) Mistry & Misener MIP & SMT Wednesday 11 th January, / 33

26 Two Dimensional Bin Packing Primal SMT UNKNOWN Return: UNKNOWN Initialise Model Solve Model Status? UNSAT Return: Optimal Solution/UNSAT SAT Set all unused bins to False. Set an extra bin to False. Mistry & Misener MIP & SMT Wednesday 11 th January, / 33

27 Two Dimensional Bin Packing Dual SMT Use unsatisfiable core UNKNOWN Return: UNKNOWN Initialise Model. All but first bins are set to False. Solve Model. Status? SAT Return: Optimal Solution UNSAT Add extra bin. Mistry & Misener MIP & SMT Wednesday 11 th January, / 33

28 Two Dimensional Bin Packing Dual SMT Fixing items in bins Mistry & Misener MIP & SMT Wednesday 11 th January, / 33

29 Two Dimensional Bin Packing Dual SMT Fixing items in bins Mistry & Misener MIP & SMT Wednesday 11 th January, / 33

30 Two Dimensional Bin Packing Dual SMT Fixing items in bins Mistry & Misener MIP & SMT Wednesday 11 th January, / 33

31 Two Dimensional Bin Packing Dual SMT Fixing items in bins Mistry & Misener MIP & SMT Wednesday 11 th January, / 33

32 Two Dimensional Bin Packing Dual SMT Fixing items in bins Mistry & Misener MIP & SMT Wednesday 11 th January, / 33

33 Two Dimensional Bin Packing Dual SMT Fixing items in bins Mistry & Misener MIP & SMT Wednesday 11 th January, / 33

34 Two Dimensional Bin Packing Dual SMT Fixing items in bins Mistry & Misener MIP & SMT Wednesday 11 th January, / 33

35 Two Dimensional Bin Packing Numerical Results Problem Set 500 problems (Lodi et al., 1999) 10 classes each class has 50 problems each problem has 20,40,60,80 or 100 items (10 of each) Each problem given a time limit of 3600s SMT solver: Z3 MILP solver: Gurobi Mistry & Misener MIP & SMT Wednesday 11 th January, / 33

36 Two Dimensional Bin Packing Numerical Results Problems Solved Class SMT-Primal SMT-Dual MILP Table: Number of problems solved to optimality within 3600s by each of the approaches. Each problem class has 50 problems. Mistry & Misener MIP & SMT Wednesday 11 th January, / 33

37 Two Dimensional Bin Packing Numerical Results (500 Bin packing problems, items) Performance Profile (Dolan and Moré, 2002) SMT-Primal SMT-Dual MILP Can solve ρ τ Best in class: Pisinger and Sigurd (2007) Constraint satisfaction + Dantzig-Wolfe decomposition Mistry & Misener MIP & SMT Wednesday 11 th January, / 33

38 Two Dimensional Bin Packing Numerical Results Pecentage Gaps Percentage gap: upper bound lower bound upper bound Class SMT MILP Mistry & Misener MIP & SMT Wednesday 11 th January, / 33

39 Scheduling Scheduling Problem Definiton Given a set of facilities each with a limit on the amount of resources available at any given time. Also, given a set of tasks where each task has release and due times and per facility resource consumption rate and processing time. What is the optimal assignment of tasks to facilities such that we are optimal according to our objective? Objectives we study Minimise cost (addition parameter - cost of assignment) Minimise makespan Minimise tardiness We study a hybrid strategy We implement the hybrid formulations of Hooker (2007) Occurs in the manufacturing and supply chain context Mistry & Misener MIP & SMT Wednesday 11 th January, / 33

40 Scheduling Mixed Integer Formulation Discrete time model Minimum cost min ijt F ij x ijt s.t. x ijt = 1 it j t T ijt c ij x ijt C i x ijt = 0, t < r j or t > d j p ij where x ijt is binary and T ijt = {t t p ij < t t}. Mistry & Misener MIP & SMT Wednesday 11 th January, / 33

41 Scheduling Logical Formulations Continuous time model Minimum cost min j f j s.t. s j r j ζ ij (s j d j p ij ) ζ ij i i i ζ ij ζ ij c ij C i (s j + p ij s j) j J j J j j ζ ij (f j = F ij ) Mistry & Misener MIP & SMT Wednesday 11 th January, / 33

42 Scheduling Logical Formulations Relation to Bin Packing Time j c ij C i C Time Link between resource constrained scheduling & bin packing Garey et al. (1976); Castro and Grossmann (2012) Mistry & Misener MIP & SMT Wednesday 11 th January, / 33

43 Scheduling Optimisation strategies Mixed-Integer Programming Problem 0 < z < 1 Satisfiability Modulo Theories tighten objective bound Subproblem 1 Subproblem 2 z = 0 z = 1 SMT optimum MIP + SMT assignment MILP SMT optimum cuts Mistry & Misener MIP & SMT Wednesday 11 th January, / 33

44 Scheduling Hybrid Scheduling Minimise cost On each iteration, add a cut that removes the current assignment or a subset (subset results in a stronger cut) Minimise makespan On each iteration, add a cut based on local makespans Minimise tardiness On each iteration, add a cut based on local tardiness Mistry & Misener MIP & SMT Wednesday 11 th January, / 33

45 Scheduling Numerical Results Experimental Setup 335 problems 1 4 classes of problems: c, de, df, e 5 problems for each (i,j) instance Each problem given a time limit of 3600s Same set of problems used and generated by Hooker (2007). We compare our findings with his MILP/CP approach. SMT solver: Z3 MILP solver: Gurobi Hooker (2007) reports that the hybrid approach is the best choice for each objective 1 Mistry & Misener MIP & SMT Wednesday 11 th January, / 33

46 Scheduling Numerical Results Minimum Cost ρ SMT MILP Hybrid Hybrid with relaxation Can solve τ Figure: Solution time performance profile for all minimum cost models. Mistry & Misener MIP & SMT Wednesday 11 th January, / 33

47 Scheduling Numerical Results Minimum Cost problems solved within τ number of problems ,000 τ = run time best run time SMT MILP Hybrid Can solve Figure: Solution time performance profile for non toy minimum cost models. Mistry & Misener MIP & SMT Wednesday 11 th January, / 33

48 Scheduling Numerical Results Minimum Makespan problems solved within τ number of problems SMT MILP 0.2 Hybrid 0.1 Can solve ,000 τ = run time best run time Figure: Solution time performance profile for non toy minimum makespan problems. Mistry & Misener MIP & SMT Wednesday 11 th January, / 33

49 Scheduling Stepping Back What computer science methods should we incorporate? Satisfiability modulo theories Software verification Potential Hybrid optimisation/logic; Optimisation explainers e.g., Mistry and Misener (2017); Lundbaek et al. (2016) Anytime algorithms Artificial intelligence Potential Improve primal solutions, Approximation algorithms Theoretical computer science Potential Heuristic solutions with performance guarantees e.g., Furman and Sahinidis (2001), Ahmed and Sahinidis (2003) Functional programming Programming languages Potential Pattern matching, parallel programming e.g., Ceccon et al. (2016), Camarda & co-workers Others: Machine learning, Data science, Internet of things, HPC Mistry & Misener MIP & SMT Wednesday 11 th January, / 33

50 Thank You Thank you for listening. Questions? Funding: EPSRC EP/M028240/1, RAEng Fellowship Mistry & Misener MIP & SMT Wednesday 11 th January, / 33

51 References I Bjørner, N. and De Moura, L. (2011). Satisfiability Modulo Theories: Introduction and Applications. Communications of the ACM, 54(9): Bockmayr, A. and Kasper, T. (1998). Branch and Infer: A Unifying Framework for Integer and Finite Domain Constraint Programming. INFORMS Journal on Computing, 10(3): Bockmayr, A. and Kasper, T. (2004). Constraint and Integer Programming: Toward a Unified Methodology, chapter Branch-and-Infer: A Framework for Combining CP and IP, pages Springer US, Boston, MA. Castro, P. M. and Grossmann, I. E. (2012). From time representation in scheduling to the solution of strip packing problems. Computers & Chemical Engineering, 44: Mistry & Misener MIP & SMT Wednesday 11 th January, / 33

52 References II Ceccon, F., Kouyialis, G., and Misener, R. (2016). Using functional programming to recognize named structure in an optimization problem: Application to pooling. AIChE Journal, 62(9): Dolan, E. D. and Moré, J. J. (2002). Benchmarking optimization software with performance profiles. Math Program, 91(2): Garey, M. R., Graham, R. L., Johnson, D. S., and Yao, A. C.-C. (1976). Resource constrained scheduling as generalized bin packing. J Combin Theory, Ser A, 21(3): Hooker, J. N. (2007). Planning and Scheduling by Logic-Based Benders Decomposition. Operations Research, 55(3): Hooker, J. N. and Ottoson, G. (2003). Logic-based Benders decomposition. Mathematical Programming, 96(1): Jain, V. and Grossmann, I. E. (2001). Algorithms for Hybrid MILP/CP Models for a Class of Optimization Problems. INFORMS Journal on Computing, 13(4): Mistry & Misener MIP & SMT Wednesday 11 th January, / 33

53 References III Lodi, A., Martello, S., and Monaci, M. (2002). Two-dimensional packing problems: A survey. European Journal of Operational Research, 141(2): Lodi, A., Martello, S., and Vigo, D. (1999). Heuristic and metaheuristic approaches for a class of two-dimensional bin packing problems. INFORMS Journal on Computing, 11(4): Lundbaek, L.-N., Callia D Iddio, A., and Huth, M. (2016). Optimizing governed blockchains for financial process authentications. arxiv preprint arxiv: Maravelias, C. T. and Grossmann, I. E. (2004). A hybrid MILP/CP decomposition approach for the continuous time scheduling of multipurpose batch plants. Computers & Chemical Engineering, 28(10): Mistry & Misener MIP & SMT Wednesday 11 th January, / 33

54 References IV Maravelias, C. T. and Sung, C. (2009). Integration of production planning and scheduling: Overview, challenges and opportunities. Computers & Chemical Engineering, 33(12): Mistry, M. and Misener, R. (2017). Integrating mixed-integer optimisation & satisfiability modulo theories: Application to scheduling. In Maravelias, Wassick, et al., editors, Computer-Aided Chemical Engineering. Pisinger, D. and Sigurd, M. (2007). Using decomposition techniques and constraint programming for solving the two-dimensional bin-packing problem. INFORMS Journal on Computing, 19(1): Trespalacios, F. and Grossmann, I. E. (2016). Symmetry breaking for generalized disjunctive programming formulation of the strip packing problem. Annals of Operations Research, pages Mistry & Misener MIP & SMT Wednesday 11 th January, / 33

Integrating Mixed-Integer Optimisation and Satisfiability Modulo Theories: Application to Scheduling

Integrating Mixed-Integer Optimisation and Satisfiability Modulo Theories: Application to Scheduling M. Mistry and R. Misener Department of Computing, Imperial College London, South Kensington Campus,