Aggregates in Recursion: Issues and Solutions

1. Motivation Basics: Answer Set Programming and Aggregates Semantics Properties Aggregates in Recursion: Issues and Solutions Wolfgang Faber Alpen-Adria-Universität Klagenfurt RuleML Webinar, 2018-05-25 Wolfgang Faber Aggregates in Recursion: Issues and Solutions

2. Motivation Basics: Answer Set Programming and Aggregates Semantics Properties Outline 1 Motivation 2 Basics: Answer Set Programming and Aggregates 3 Semantics Stratiﬁed Aggregates Unstratiﬁed Aggregates 4 Properties Coincidence Results Complexity Results Wolfgang Faber Aggregates in Recursion: Issues and Solutions

3. Motivation Basics: Answer Set Programming and Aggregates Semantics Properties Motivation: Aggregates Aggregates facilitate problem representation Standard in database query languages For example: SELECT company.name FROM company,employee WHERE company.id = employee.cid AND COUNT(employee.id) < 10 Wolfgang Faber Aggregates in Recursion: Issues and Solutions

4. Motivation Basics: Answer Set Programming and Aggregates Semantics Properties Motivation: Aggregates and Logic Programming In this talk: Combination of Logic Programming and Aggregates For example: SELECT company.name FROM company,employee WHERE company.id = employee.cid AND COUNT(employee.id) < 10 using Logic Programming (here ASP/Datalog): result(N) :- company(N, Y), #count{X : employee(X, Y)} < 10. Does this look innocent? Wolfgang Faber Aggregates in Recursion: Issues and Solutions

5. Motivation Basics: Answer Set Programming and Aggregates Semantics Properties Recursion A main feature of Logic Programming is recursion Example: tc(A, B) :- p(A, B). tc(A, B) :- p(A, X), tc(X, B). deﬁnes the transitive closure of p Supported also in SQL Recent (since SQL-99) Not particularly well-known Common Table Expressions (CTE) Wolfgang Faber Aggregates in Recursion: Issues and Solutions

6. Motivation Basics: Answer Set Programming and Aggregates Semantics Properties Recursion and Aggregates Combination of recursion and aggregates? Explicitly forbidden in SQL! Meaning of p(a) :- #count{X : p(X)} > 0. Meaning of p(a) :- #count{X : p(X)} < 1. Meaning of p(1) :- #avg{X : p(X)}! = 1. p(−1) :- #avg{X : p(X)}! = 1. Wolfgang Faber Aggregates in Recursion: Issues and Solutions

7. Motivation Basics: Answer Set Programming and Aggregates Semantics Properties ASP Semantics Herbrand interpretations Reduct for interpretation I: 1 Delete rules whose negative body is true in I. 2 Delete negative body from all other rules. Interpretations I which are minimal models of the reduct for I are answer sets. [Gelfond, Lifschitz 1988] (nondisjunctive) [Przymusinski 1991] (disjunctive) [Gelfond, Lifschitz 1991] (disjunctive, 2 kinds of negation) Wolfgang Faber Aggregates in Recursion: Issues and Solutions

8. Motivation Basics: Answer Set Programming and Aggregates Semantics Properties ASP Semantics Herbrand interpretations Reduct for interpretation I: 1 Delete rules whose negative body is true in I. 2 Delete negative body from all other rules. Interpretations I which are minimal models of the reduct for I are answer sets. [Gelfond, Lifschitz 1988] (nondisjunctive) [Przymusinski 1991] (disjunctive) [Gelfond, Lifschitz 1991] (disjunctive, 2 kinds of negation) Wolfgang Faber Aggregates in Recursion: Issues and Solutions

9. Motivation Basics: Answer Set Programming and Aggregates Semantics Properties Aggregates Aggregate Functions: Functions over ground term (multi)sets (Multi)sets speciﬁed as {A, B : Conj} or { c, d : Conj , . . .} Evaluate Conj w.r.t. an interpretation Aggregate Atoms: Aggregate Function plus comparison Important: Aggregate atoms depend on truth values of a set of standard atoms! Wolfgang Faber Aggregates in Recursion: Issues and Solutions

13. Motivation Basics: Answer Set Programming and Aggregates Semantics Properties Stratified Aggregates Unstratified Aggregates Outline 1 Motivation 2 Basics: Answer Set Programming and Aggregates 3 Semantics Stratified Aggregates Unstratified Aggregates 4 Properties Coincidence Results Complexity Results Wolfgang Faber Aggregates in Recursion: Issues and Solutions

14. Motivation Basics: Answer Set Programming and Aggregates Semantics Properties Stratified Aggregates Unstratified Aggregates Aggregate Stratification Definition A program P is stratified on an aggregate atom A if there exists a level mapping || || from its predicates to ordinals, such that for each rule and for each of its head atoms a the following holds: 1 For each predicate b of standard body literals: ||b|| ≤ ||a||, 2 for each predicate b inside an aggregate body atom: ||b|| < ||a||, and 3 for each predicate b in the head: ||b|| = ||a||. Wolfgang Faber Aggregates in Recursion: Issues and Solutions

15. Motivation Basics: Answer Set Programming and Aggregates Semantics Properties Stratified Aggregates Unstratified Aggregates Aggregate Stratification Note: Stratification is relative to a program Unstratified aggregate atoms occur recursively Example a :- #count{ t : b } > 0. is stratified on the aggregate atom. a :- #count{ t : b } > 0. b :- a. is not stratified on the aggregate atom. Wolfgang Faber Aggregates in Recursion: Issues and Solutions

16. Motivation Basics: Answer Set Programming and Aggregates Semantics Properties Stratified Aggregates Unstratified Aggregates Answer Sets for Aggregate-stratified Programs Basic Idea: Treat aggregate atoms like negative literals Reduct: 1 Delete rules containing unsatisfied aggregates and negative literals 2 Delete aggregates and negative literals from all other rules [Kemp, Stuckey 1991], [Gelfond 2002], [Dell’Armi, F., Ielpa, Leone, Pfeifer 2003] Many programs are aggregate stratified Use of aggregates often yields computational advantages But: Not all programs are aggregate stratified Wolfgang Faber Aggregates in Recursion: Issues and Solutions

19. Motivation Basics: Answer Set Programming and Aggregates Semantics Properties Stratified Aggregates Unstratified Aggregates Outline 1 Motivation 2 Basics: Answer Set Programming and Aggregates 3 Semantics Stratified Aggregates Unstratified Aggregates 4 Properties Coincidence Results Complexity Results Wolfgang Faber Aggregates in Recursion: Issues and Solutions

20. Motivation Basics: Answer Set Programming and Aggregates Semantics Properties Stratified Aggregates Unstratified Aggregates Unstratification What happens when we consider unstratified aggregates? Can we just keep the simple semantic definition? Wolfgang Faber Aggregates in Recursion: Issues and Solutions

21. Motivation Basics: Answer Set Programming and Aggregates Semantics Properties Stratiﬁed Aggregates Unstratiﬁed Aggregates Company Control Input: Set of companies and shares companies hold of other companies. Output: Companies controlled (direct or indirect shares > 50%) by other companies Encoding from the literature: controlsStk(C1, C1, C2, P) :- ownsStk(C1, C2, P). controlsStk(C1, C2, C3, P) :- controls(C1, C2), ownsStk(C2, C3, P). controls(C1, C3) :- company(C1), company(C3), #sum{P, C2 : controlsStk(C1, C2, C3, P)} > 50. Wolfgang Faber Aggregates in Recursion: Issues and Solutions

22. Motivation Basics: Answer Set Programming and Aggregates Semantics Properties Stratiﬁed Aggregates Unstratiﬁed Aggregates Company Control, Instance 1 Example (Company Control, Instance 1) 30% 40% 60% 35% 15% a c d b {controlsStk(a,a,b,60), controlsStk(a,a,c,30), controlsStk(a,a,d,15), controlsStk(b,b,d,40), controlsStk(c,c,d,35), controlsStk(a,b,d,40), controls(a,b), controls(a,d)} Wolfgang Faber Aggregates in Recursion: Issues and Solutions

24. Motivation Basics: Answer Set Programming and Aggregates Semantics Properties Stratiﬁed Aggregates Unstratiﬁed Aggregates Company Control, Instance 2 Example (Company Control, Instance 2) a cb 40% 40% 20% 20% {controlsStk(a,a,b,40), controlsStk(a,a,c,40), controlsStk(b,b,c,20), controlsStk(c,c,b,20)} But also: {controlsStk(a,b,c,20), controlsStk(a,c,b,20), controls(a,b), controls(a,c)} Wolfgang Faber Aggregates in Recursion: Issues and Solutions

27. Motivation Basics: Answer Set Programming and Aggregates Semantics Properties Stratiﬁed Aggregates Unstratiﬁed Aggregates Essence Example a :- #count{ t : a } < 1. No answer sets. a :- #count{ t : a } > 0. Answer sets: ∅, {a}? #count{ t : a } < 1 behaves like not a #count{ t : a } > 0 behaves like a ⇒ aggregates should not be treated like negative literals, but also not like positive literals Wolfgang Faber Aggregates in Recursion: Issues and Solutions

28. Motivation Basics: Answer Set Programming and Aggregates Semantics Properties Stratiﬁed Aggregates Unstratiﬁed Aggregates Essence Example a :- #count{ t : a } < 1. No answer sets. a :- #count{ t : a } > 0. Answer sets: ∅, {a}? #count{ t : a } < 1 behaves like not a #count{ t : a } > 0 behaves like a ⇒ aggregates should not be treated like negative literals, but also not like positive literals Wolfgang Faber Aggregates in Recursion: Issues and Solutions

29. Motivation Basics: Answer Set Programming and Aggregates Semantics Properties Stratiﬁed Aggregates Unstratiﬁed Aggregates Monotonicity and Antimonotonicity Monotone Literals: truth for interpretation I implies truth for all J ⊇ I Antimonotone Literals: truth for interpretation J implies truth for all I ⊆ J Nonmonotone Literals: neither monotone nor antimonotone Wolfgang Faber Aggregates in Recursion: Issues and Solutions

30. Motivation Basics: Answer Set Programming and Aggregates Semantics Properties Stratiﬁed Aggregates Unstratiﬁed Aggregates Monotonicity: Examples #count{. . .} ≥ 1 is monotone #count{. . .} < 1 is antimonotone #avg{. . .} < 3 is nonmonotone Positive standard literals are monotone Negative standard literals are antimonotone Wolfgang Faber Aggregates in Recursion: Issues and Solutions

31. Motivation Basics: Answer Set Programming and Aggregates Semantics Properties Stratified Aggregates Unstratified Aggregates FLP Semantics: Novel Reduct Definition Definition of reduct according to [F.,Leone,Pfeifer 2004, F., Leone, Pfeifer 2011]: Delete rules with a false body literal. That’s it! Answer Set: Subset-minimal model of the reduct Theorem For aggregate-free programs, answer sets under this definition coincide with the ones defined in [Gelfond, Lifschitz 1991]. (Mostly) equivalent semantics defined in [Ferraris 2005], [Ferraris 2011]. Wolfgang Faber Aggregates in Recursion: Issues and Solutions

36. Motivation Basics: Answer Set Programming and Aggregates Semantics Properties Stratiﬁed Aggregates Unstratiﬁed Aggregates Company Control, Instance 2 Example (Company Control, Instance 2) a cb 40% 40% 20% 20% {controlsStk(a,a,b,40), controlsStk(a,a,c,40), controlsStk(b,b,c,20), controlsStk(c,c,b,20)} Only this answer set. Wolfgang Faber Aggregates in Recursion: Issues and Solutions

37. Motivation Basics: Answer Set Programming and Aggregates Semantics Properties Stratiﬁed Aggregates Unstratiﬁed Aggregates Company Control, Instance 2 Example (Company Control, Instance 2) a cb 40% 40% 20% 20% {controlsStk(a,a,b,40), controlsStk(a,a,c,40), controlsStk(b,b,c,20), controlsStk(c,c,b,20)} Only this answer set. Wolfgang Faber Aggregates in Recursion: Issues and Solutions

38. Motivation Basics: Answer Set Programming and Aggregates Semantics Properties Stratiﬁed Aggregates Unstratiﬁed Aggregates Small examples Example a :- #count{ t : a } < 1. No answer sets. a :- #count{ t : a } > 0. Answer sets: only ∅ Wolfgang Faber Aggregates in Recursion: Issues and Solutions

39. Motivation Basics: Answer Set Programming and Aggregates Semantics Properties Stratified Aggregates Unstratified Aggregates PSP Semantics Alternative semantics for unstratified aggregates: [Pelov 2004], [Son, Pontelli 2007], [Shen, Wang 2012] Definitions use different operator-based techniques Evaluate aggregates for a pair of interpretations (I, J) |= A iff K |= A for all I ⊆ K ⊆ J. PSP answer sets are those I that are fixpoints of KI Π ⇑ ∅ Operator KI Π(X) collects heads of rules for which (X, I) |= A for all body atoms Wolfgang Faber Aggregates in Recursion: Issues and Solutions

40. Motivation Basics: Answer Set Programming and Aggregates Semantics Properties Stratiﬁed Aggregates Unstratiﬁed Aggregates Small examples: PSP Example a :- #count{ t : a } < 1. No answer sets. K∅ Π(∅) = {a} K∅ Π({a}) = {a} K {a} Π (∅) = ∅ a :- #count{ t : a } > 0. Answer sets: only ∅ K∅ Π(∅) = ∅ K {a} Π (∅) = ∅ Wolfgang Faber Aggregates in Recursion: Issues and Solutions

41. Motivation Basics: Answer Set Programming and Aggregates Semantics Properties Coincidence Results Complexity Results Outline 1 Motivation 2 Basics: Answer Set Programming and Aggregates 3 Semantics Stratiﬁed Aggregates Unstratiﬁed Aggregates 4 Properties Coincidence Results Complexity Results Wolfgang Faber Aggregates in Recursion: Issues and Solutions

42. Motivation Basics: Answer Set Programming and Aggregates Semantics Properties Coincidence Results Complexity Results FLP and PSP Small examples could suggest that they coincide. But this is not true in general. Example p(1) :- #avg{X : p(X)} ≥ 0. p(1) :- p(−1). p(−1) :- p(1). FLP answer sets: {p(1), p(−1)} PSP answer sets: none In general: each PSP answer set is also FLP, but not necessarily vice versa. Classes on which the semantics coincide? Wolfgang Faber Aggregates in Recursion: Issues and Solutions

43. Motivation Basics: Answer Set Programming and Aggregates Semantics Properties Coincidence Results Complexity Results Aggregate-stratiﬁed Programs Easy observation: Theorem FLP and PSP semantics coincide on aggregate-stratiﬁed programs. Wolfgang Faber Aggregates in Recursion: Issues and Solutions

44. Motivation Basics: Answer Set Programming and Aggregates Semantics Properties Coincidence Results Complexity Results Monotone, antimonotone, convex Wolfgang Faber Aggregates in Recursion: Issues and Solutions {a, b, c, d} {a, b, c} {a, b, d} {a, c, d} {b, c, d} {a, b} {a, c} {a, d} {b, c} {b, d} {c, d} {a} {b} {c} {d} ∅ 1 ≤ #count{a, b, c, d} S is monotone if I |= S ∧ J |= S =⇒ K |= S ∀K ∈ ↑I ∩ ↓J

45. Motivation Basics: Answer Set Programming and Aggregates Semantics Properties Coincidence Results Complexity Results Monotone, antimonotone, convex Wolfgang Faber Aggregates in Recursion: Issues and Solutions {a, b, c, d} {a, b, c} {a, b, d} {a, c, d} {b, c, d} {a, b} {a, c} {a, d} {b, c} {b, d} {c, d} {a} {b} {c} {d} ∅ 1 ≤ #count{a, b, c, d} S is monotone if I |= S ∧ J |= S =⇒ K |= S ∀K ∈ ↑I ∩ ↓J

46. Motivation Basics: Answer Set Programming and Aggregates Semantics Properties Coincidence Results Complexity Results Monotone, antimonotone, convex Wolfgang Faber Aggregates in Recursion: Issues and Solutions {a, b, c, d} {a, b, c} {a, b, d} {a, c, d} {b, c, d} {a, b} {a, c} {a, d} {b, c} {b, d} {c, d} {a} {b} {c} {d} ∅ #count{a, b, c, d} ≤ 3 S is antimonotone if I |= S ∧ J |= S =⇒ K |= S ∀K ∈ ↑I ∩ ↓J

47. Motivation Basics: Answer Set Programming and Aggregates Semantics Properties Coincidence Results Complexity Results Monotone, antimonotone, convex Wolfgang Faber Aggregates in Recursion: Issues and Solutions {a, b, c, d} {a, b, c} {a, b, d} {a, c, d} {b, c, d} {a, b} {a, c} {a, d} {b, c} {b, d} {c, d} {a} {b} {c} {d} ∅ #count{a, b, c, d} ≤ 3 S is antimonotone if I |= S ∧ J |= S =⇒ K |= S ∀K ∈ ↑I ∩ ↓J

48. Motivation Basics: Answer Set Programming and Aggregates Semantics Properties Coincidence Results Complexity Results Monotone, antimonotone, convex Wolfgang Faber Aggregates in Recursion: Issues and Solutions {a, b, c, d} {a, b, c} {a, b, d} {a, c, d} {b, c, d} {a, b} {a, c} {a, d} {b, c} {b, d} {c, d} {a} {b} {c} {d} ∅ 1 ≤ #count{a, b, c, d} ≤ 3 S is convex if I |= S ∧ J |= S =⇒ K |= S ∀K ∈ ↑I ∩ ↓J

49. Motivation Basics: Answer Set Programming and Aggregates Semantics Properties Coincidence Results Complexity Results Monotone, antimonotone, convex Wolfgang Faber Aggregates in Recursion: Issues and Solutions {a, b, c, d} {a, b, c} {a, b, d} {a, c, d} {b, c, d} {a, b} {a, c} {a, d} {b, c} {b, d} {c, d} {a} {b} {c} {d} ∅ 1 ≤ #count{a, b, c, d} ≤ 3 S is convex if I |= S ∧ J |= S =⇒ K |= S ∀K ∈ ↑I ∩ ↓J

50. Motivation Basics: Answer Set Programming and Aggregates Semantics Properties Coincidence Results Complexity Results Convex Programs Implicit in [Liu, Truszczy´nski 2006]: Theorem FLP and PSP semantics coincide on programs containing only convex aggregates. Corollary FLP and PSP semantics coincide on programs containing only monotone and antimonotone aggregates. Wolfgang Faber Aggregates in Recursion: Issues and Solutions

51. Motivation Basics: Answer Set Programming and Aggregates Semantics Properties Coincidence Results Complexity Results Outline 1 Motivation 2 Basics: Answer Set Programming and Aggregates 3 Semantics Stratiﬁed Aggregates Unstratiﬁed Aggregates 4 Properties Coincidence Results Complexity Results Wolfgang Faber Aggregates in Recursion: Issues and Solutions

52. Motivation Basics: Answer Set Programming and Aggregates Semantics Properties Coincidence Results Complexity Results Problem Cautious reasoning over variable-free programs and polynomial-time computable aggregate functions. Input: A ground program P and a ground standard atom A. Output: Is A true in all FLP answer sets of P? Similar results for related problems (answer set existence, brave reasoning). Wolfgang Faber Aggregates in Recursion: Issues and Solutions

53. Motivation Basics: Answer Set Programming and Aggregates Semantics Properties Coincidence Results Complexity Results Problem Cautious reasoning over variable-free programs and polynomial-time computable aggregate functions. Input: A ground program P and a ground standard atom A. Output: Is A true in all FLP answer sets of P? Similar results for related problems (answer set existence, brave reasoning). Wolfgang Faber Aggregates in Recursion: Issues and Solutions

54. Motivation Basics: Answer Set Programming and Aggregates Semantics Properties Coincidence Results Complexity Results Complexity of Cautious Reasoning {} {not} {∨} {not, ∨} {M} P co-NP co-NP ΠP 2 {S} P co-NP ΠP 2 ΠP 2 {C} co-NP co-NP ΠP 2 ΠP 2 {N} ΠP 2 ΠP 2 ΠP 2 ΠP 2 {M}: monotone aggregates {S}: stratiﬁed aggregates {C}: convex aggregates {N}: non-convex aggregates Wolfgang Faber Aggregates in Recursion: Issues and Solutions

55. Motivation Basics: Answer Set Programming and Aggregates Semantics Properties Coincidence Results Complexity Results Complexity: FLP and PSP Results for PSP: analogous to FLP Slight difference for non-convex aggregates [Alviano, F. 2013]: One non-convex aggregate is sufﬁcient to express any problem in ΠP 2 with FLP Arbitrarily many are needed to do the same with PSP Wolfgang Faber Aggregates in Recursion: Issues and Solutions

56. Motivation Basics: Answer Set Programming and Aggregates Semantics Properties Coincidence Results Complexity Results Only Aggregates? All of the results presented here also apply to other extensions of ASP that have constructs that evaluate truth on sets of basic atoms, for example: Abstract Constraint Atoms (U, V) [Marek, Remmel 2004] HEX programs [Eiter et al. 2005] Nested Expressions [Lifschitz et al. 1999] Generalized Quantiﬁers [Lindström 1966] Cardinality and Weight Constraints [Simons 2000] Wolfgang Faber Aggregates in Recursion: Issues and Solutions

57. Motivation Basics: Answer Set Programming and Aggregates Semantics Properties Coincidence Results Complexity Results Summary Aggregates in ASP FLP and PSP Semantics Properties: Coincidence and Complexity Wolfgang Faber Aggregates in Recursion: Issues and Solutions

58. Motivation Basics: Answer Set Programming and Aggregates Semantics Properties Coincidence Results Complexity Results More Beyond FLP and PSP? [Alviano, F. 2015] Example a :- #count{a, b}! = 1. b :- #count{a, b}! = 1. No FLP, no PSP answer sets! Unintuitive? Reduce programs with non-convex aggregates to programs with monotone aggregates in a compact way. [Alviano, F., Gebser 2015] System support: results in [Alviano, F., Gebser 2015] allow for non-convex aggregates in gringo ≥ 4.5. Wolfgang Faber Aggregates in Recursion: Issues and Solutions

Aggregates in Recursion: Issues and Solutions

More Related Content

Similar to Aggregates in Recursion: Issues and Solutions (20)

More from RuleML (20)

Recently uploaded (20)

Aggregates in Recursion: Issues and Solutions