1. Happy Mittal (Joint work with Prasoon Goyal, Parag Singla and Vibhav Gogate) happy.mittal@cse.iitd.ac.in IIT Delhi New Rules for Domain Independent Lifted MAP Inference

2. Overview Introduction Motivation First rule Second rule Experiments Conclusion and Future work

3. Markov Logic Networks (MLN) Real world problems characterized by ◦ Entities and Relationships & Uncertain Behavior Relational Models : Horn clauses, SQL, first-order logic Statistical Models: Markov networks, Bayesian networks Markov Logic [Richardson & Domingos 06] Weighted First Order Formulas Smoking leads to Cancer Friendship leads to Similar Smoking habits

4. Example: Friends & Smokers S(A)S(B)F(A,A)F(A,B)F(B,A)F(B,B)P 000000exp(8.0) 100100exp(6.0).............. 111111exp(8.0) Ground formulas Weight of formula iNo. of true groundings of formula i in x

5. MAP/MPE Inference This is just the weighted MaxSAT problem Use weighted SAT solver (e.g., MaxWalkSAT [Kautz et al. 97] ) S(A)S(B)F(A,A)F(A,B)F(B,A)F(B,B)P 000000exp(8.0) 100100exp(6.0).............. 111111exp(8.0)

6. Overview Introduction Motivation Notations and preliminaries First rule Second rule Experiments Conclusion

7. MAP/MPE Inference Naïve way : Solve on fully grounded theory Exponential in size of domain of predicates. Exploit symmetries in the theory (Lifted MAP inference) How ?

8. MAP/MPE Inference All are identical up to constant renaming. Solve one, get solutions for others also. But… Some Groundings are lost. Why ? (pred,pos) pairs (P,1) and (Q,1) are related Related (pred,pos) pairs with different variables in same clause => Problem … …

9. Overview Introduction Motivation Notations and preliminaries First rule Second rule Experiments Conclusion and Future work

10. First rule for lifted MAP inference Split over single occurrence (pred,pos) pairs MAP inference over M can be reduced to MAP inference over any of M1, M2, M3. M1 M2M3 Split over (Q,2)

11. First rule for lifted MAP inference M1 M2M3 Split over (R,1) MAP inference in single occurrence MLN is domain independent.

12. Overview Introduction Motivation Notations and preliminaries First rule Second rule Experiments Conclusion and Future work

13. Extreme Assignment Theorem : If Predicate P is single occurrence in M, then P has an extreme assignment in X MAP M1 M2M3

14. Exploiting Extremes Consider MLN : W1 : P(X) => Q(X,Y) W2 : Q(X,Y) ^ Q(Y,Z) => Q(X,Z) Pairs (Q,1) and (Q,2) not single occurrence But in MLN : W1 : P(X) => Q(X,Y) W2 : Q(X,Y) ^ Q(Y,Z) => Q(X,Z) (Q,1) and (Q,2) are single occurrence Hence solution on extremes. Q(X,Y) ^ Q(Y,Z) => Q(X,Z) is tautology at extremes. True ^ True => True False ^ False => False

15. Second rule for Lifting MAP Second Rule : If MLN is single occurrence after removing tautology at extremes, then MAP inference in original MLN is domain independent.

16. Overview Introduction Motivation Notations and preliminaries First rule Second rule Experiments Conclusion and Future Work

17. Experiments Compared our performance with Purely grounded version of benchmark MLNs Sarkhel et al. [2014]‘s non shared MLN approach Used ILP based solver Gurobi as base solver. Notation : GRB : purely grounded version NSLGRB : Sarkhel et al. [2014]’s non shared MLN approach SOLGRB : our approach (single occurrence lifted GRB)

18. Datasets and Methodology Used 2 benchmark MLNs : Data set No. of form ulas Single Occurr ence Taut. At extreme Example formulas IE7No (Mix)NoToken(t,p,c)=>InField(p,f,c) !Equal(f1,f2) ^ InField(p,f1,c) => !InField(p,f2,c) etc FS5No (Mix)YesS(x) => C(x) F(x,y) ^ F(y,z) => F(x,z) etc

19. Results Time to reach optimal IEFS

20. Results Cost of unsatisfied formulae with varying time (fixed domain size 500) IEFS

21. Results Size of ground theory with varying domain size. IEFS

22. Overview Introduction Motivation Notations and preliminaries First rule Second rule Experiments Conclusion and Future work

23. Presented two new rules for lifting MAP inference over MLN. Applicable to large classes of MLNs MAP inference becomes domain independent for single occurrence MLN. Can handle transitivity rules. Can we combine our new rules with existing techniques ? Incorporate evidence. Lift any structured CSP (in general any structured LP)

24. THANK YOU