Implementation of an αbb-type underestimator in the SGO-algorithm

Transcription

1 Implementation of an αbb-type underestimator in the SGO-algorithm Process Design & Systems Engineering November 3, 2010

2 Refining without branching Could the αbb underestimator be used without an explicit branching framework? (cf. the SGO algorithm)

3 Refining without branching Could the αbb underestimator be used without an explicit branching framework? (cf. the SGO algorithm) We developed a convex formulation that handles breakpoints with binary variables instead of direct branching

4 Refining without branching Could the αbb underestimator be used without an explicit branching framework? (cf. the SGO algorithm) We developed a convex formulation that handles breakpoints with binary variables instead of direct branching Why?

5 Refining without branching Could the αbb underestimator be used without an explicit branching framework? (cf. the SGO algorithm) We developed a convex formulation that handles breakpoints with binary variables instead of direct branching Why? It can readily be integrated with the SGO algorithm It could turn out to be especially well-suited for some types of mixed-integer problems As a convex reformulation it could be of interest in automated reformulation procedures

6 Refining without branching

7 Refining without branching α 1 = 1, α 2 = 0.6, α 3 = 0

8 Refining without branching α = max k α k

9 Refining without branching The underestimation error can be described as the difference of a parabola and a piecewise linear function

10 Refining without branching Our formulation: (1D for clarity) f (x) 0 (1)

11 Refining without branching Our formulation: (1D for clarity) f (x) + αx 2 0 (1)

12 Refining without branching Our formulation: (1D for clarity) f (x) + αx 2 W 0 (1)

13 Refining without branching Our formulation: (1D for clarity) f (x) + αx 2 W 0 (1) W = αx 2 (2)

14 Refining without branching Our formulation: (1D for clarity) f (x) + αx 2 W 0 (1) W = αx 2 (2) Overestimating W will relax the feasible domain, we replace W with a piecewise linear function where K Ŵ = A k b k + (B k A k )s k (3) k=1 2 A k = αx k 2 B k = αx k (x k and x k denote the interval endpoints)

15 Refining without branching The b k are binary variables, s k are continuous and nonnegative. K Ŵ = A k b k + (B k A k )s k k=1

16 Refining without branching The b k are binary variables, s k are continuous and nonnegative. K Ŵ = A k b k + (B k A k )s k k=1 We relate x to b k and s k with the constraints K b k = 1 k=1 s k b k, k K x = x k b k + (x k x k )s k k=1

17 Refining without branching The b k are binary variables, s k are continuous and nonnegative. K Ŵ = A k b k + (B k A k )s k k=1 We relate x to b k and s k with the constraints K b k = 1 k=1 s k b k, k K x = x k b k + (x k x k )s k k=1 Every constraint is convex and the feasible set is relaxed

18 In two dimensions An example: f (x, y) = sin(x + y) + x cos y 1 x 7, 1 y

19 Underestimator - no breakpoints 6 x y 4 2

20 Underestimator - 1 breakpoint 6 x y 4 2

21 Underestimator breakpoints 6 x y 4 2

22 Underestimator breakpoints 6 x y 4 2

23 Underestimator breakpoints 6 x y 4 2

24 Underestimator breakpoints 8 y Anders Skja l x

25 Constraint feasibility

26 Constraint feasibility breakpoints

27 Constraint feasibility breakpoints

28 Constraint feasibility breakpoints

29 About convergence The largest underestimation error in a subdomain depends only on the value of α i, i = 1,... n and the size of the subdomain: (1D) ( max Ŵ αx 2 = max α(x x xk x k k)(x x k ) = α x [x k,x k ] x [x k,x k ] 2 An ɛ precision is guaranteed if the width of the interval 4ɛ x k x k α. The algorithm will converge ) 2

30 Challenges & Ideas The subproblems grow as we add breakpoints, the branching is hidden in the complexity of the convex MINLPs

31 Challenges & Ideas The subproblems grow as we add breakpoints, the branching is hidden in the complexity of the convex MINLPs The increase in complexity depends on the number of breakpoints, not directly on the number of constraints

32 Challenges & Ideas The subproblems grow as we add breakpoints, the branching is hidden in the complexity of the convex MINLPs The increase in complexity depends on the number of breakpoints, not directly on the number of constraints Bound reductions are only partially applicable

33 Challenges & Ideas The subproblems grow as we add breakpoints, the branching is hidden in the complexity of the convex MINLPs The increase in complexity depends on the number of breakpoints, not directly on the number of constraints Bound reductions are only partially applicable We get less information about subdomains as compared to branch-and-bound

34 Challenges & Ideas The subproblems grow as we add breakpoints, the branching is hidden in the complexity of the convex MINLPs The increase in complexity depends on the number of breakpoints, not directly on the number of constraints Bound reductions are only partially applicable We get less information about subdomains as compared to branch-and-bound A type of minor breakpoint halving every interval can be introduced without too much cost

35 Integration with SGO Preprocessing step: Transformations obtained by solving an MILP problem The nonconvex terms are replaced by convex underestimators; the PLFs are initialized The convexified and overestimated problem is solved using a convex MINLP solver Update approximations no Are the termination criteria fulfilled? yes Optimal solution

36 Integration with SGO Preprocessing step: Transformations obtained by solving an MILP problem The nonconvex terms are replaced by convex underestimators; the PLFs are initialized The convexified and overestimated problem is solved using a convex MINLP solver Update approximations no Are the termination criteria fulfilled? yes Optimal solution

37 References C.S. Adjiman, S. Dallwig, C.A. Floudas, and A. Neumaier. A global optimization method, αbb, for general twice-differentiable constrained NLPs I. Computers & Chemical Engineering, 22(9): , C.A. Floudas. Deterministic Global Optimization. Kluwer Academic Publishers, A. Lundell, J. Westerlund, and T. Westerlund. Some transformation techniques with applications in global optimization. Journal of Global Optimization, 43(2): , A. Lundell and T. Westerlund. Convex underestimation strategies for signomial functions. Optimization Methods and Software, 24: , 2009.

38 Thank you

39 C 2 constraints We are interested in handling constraints f (x) 0 where f C 2, i.e. f is twice continuously differentiable.

40 C 2 constraints We are interested in handling constraints f (x) 0 where f C 2, i.e. f is twice continuously differentiable. In addition C 2 objective functions can be handled by rewriting min f (x) as min µ s.t. f (x) µ 0

41 The αbb underestimator A C 2 function f on the domain [x L, x U ] R can always be convexified by adding a parabola p(x) = α(x x L )(x x U ) with a large enough α. 4 2 function parabola underest x

42 The αbb underestimator A C 2 function f on the domain [x L, x U ] R can always be convexified by adding a parabola p(x) = α(x x L )(x x U ) with a large enough α. 4 2 function parabola underest x

43 The αbb underestimator This convex underestimator can be extended to multiple dimensions. Let f be a C 2 function on R n. For a large enough α the function g = f + q where is convex. q(x 1,..., x n ) = α n (x i xi L )(x i xi U ) i=1

44 The αbb underestimator This convex underestimator can be extended to multiple dimensions. Let f be a C 2 function on R n. For a large enough α the function g = f + q where q(x 1,..., x n ) = α n (x i xi L )(x i xi U ) is convex. Tighter underestimators can be found by letting α depend on i i=1 q(x 1,..., x n ) = n α i (x i xi L )(x i xi U ) i=1

45 The αbb underestimator How large must we choose α i? One dimension: g is convex if g = f + 2α 0

46 The αbb underestimator How large must we choose α i? One dimension: g is convex if g = f + 2α 0 In general: g is convex if the Hessian matrix H g is positive semi-definite

47 The αbb underestimator How large must we choose α i? One dimension: g is convex if g = f + 2α 0 In general: g is convex if the Hessian matrix H g is positive semi-definite H g = H f + 2 diag(α i ) = 2 f x 1 x 1 2 f x 1 x n α f x n x 1 2 f x n x n α n

48 The αbb underestimator How large must we choose α i? One dimension: g is convex if g = f + 2α 0 In general: g is convex if the Hessian matrix H g is positive semi-definite H g = H f + 2 diag(α i ) = 2 f x 1 x 1 2 f x 1 x n α f x n x 1 2 f x n x n Choose α i such that the eigenvalues of H g are nonnegative on the relevant domain α n

49 The αbb underestimator The α i are calculated (e.g.) by utilizing interval arithmetic and Gershgorin s circle theorem

50 The αbb underestimator The α i are calculated (e.g.) by utilizing interval arithmetic and Gershgorin s circle theorem Smaller valid choices usually exist, but finding the optimal α i is in general as hard as the optimization problem itself

51 The αbb underestimator The α i are calculated (e.g.) by utilizing interval arithmetic and Gershgorin s circle theorem Smaller valid choices usually exist, but finding the optimal α i is in general as hard as the optimization problem itself Branch-and-bound methods can be used to solve the original problem

52 Branch-and-bound search Branch to split a domain in two and get tighter underestimators

53 Branch-and-bound search Branch to split a domain in two and get tighter underestimators Any feasible solution to the original problem gives an upper bound on the optimal objective value

54 Branch-and-bound search Branch to split a domain in two and get tighter underestimators Any feasible solution to the original problem gives an upper bound on the optimal objective value Any branch with a lower bound greater than the best found upper bound is fathomed (cut)

55 Branch-and-bound search Branch to split a domain in two and get tighter underestimators Any feasible solution to the original problem gives an upper bound on the optimal objective value Any branch with a lower bound greater than the best found upper bound is fathomed (cut) A number of techniques are used to speed up the search, e.g. bound reduction