A Talk about the E deduction server

1. deduction as a service
Mohamed Bassem
July 28, 2015
German University in Cairo

2. table of contents
1. What is E?
2. Large Axiom Sets
3. Problem
4. Deduction Server
5. Results
6. Future Work
7. Conclusion
2

3. what is e?

4. what is e?
E is a theorem prover that takes a set of axioms and tries to derive a
formal proof to a given conjuncture
4

5. what is e?
E is a theorem prover that takes a set of axioms and tries to derive a
formal proof to a given conjuncture
Example:
4

6. large axiom sets

7. large axiom sets
• OpenCYC library tries to collect the human’s common sense
knowledge into a comprehensive library with the aim of allowing
AI applications to perfom human-like reasoning[1]
6

8. large axiom sets
• OpenCYC library tries to collect the human’s common sense
knowledge into a comprehensive library with the aim of allowing
AI applications to perfom human-like reasoning[1]
• Up to 500MB of data and millions of axioms
6

9. large axiom sets
• OpenCYC library tries to collect the human’s common sense
knowledge into a comprehensive library with the aim of allowing
AI applications to perfom human-like reasoning[1]
• Up to 500MB of data and millions of axioms
• Other huge axiom sets are extracted from libraries such as
SUMO and MIZAR
6

10. problem

11. problem
Usually queries only reRunning multiple queries against the same
knowledge base requires re-parsing the whole set
8

12. problem
No formal way of communicating with E
9

13. deduction server

14. revisiting the problem
11

15. deduction server: sessions
12

16. deduction server: server client architecture
13

17. deduction server: pruning
14

18. deduction server: parallelization
15

19. deduction server: interaction protocol
16

20. deduction server
Live Demo!
17

21. results

22. benchmarking
• Tests were performed on a virtualized server with 4 2.6 GHz Intel
CPUs with 8 threads
19

23. benchmarking
• Tests were performed on a virtualized server with 4 2.6 GHz Intel
CPUs with 8 threads
• Chosen TPTP Problems:
19

24. benchmarking
• Tests were performed on a virtualized server with 4 2.6 GHz Intel
CPUs with 8 threads
• Chosen TPTP Problems:
• 51 Problems
19

25. benchmarking
• Tests were performed on a virtualized server with 4 2.6 GHz Intel
CPUs with 8 threads
• Chosen TPTP Problems:
• 51 Problems
• 479 Megabytes on disk
19

26. benchmarking
• Tests were performed on a virtualized server with 4 2.6 GHz Intel
CPUs with 8 threads
• Chosen TPTP Problems:
• 51 Problems
• 479 Megabytes on disk
• 3,341,984 Formulae
19

27. benchmarking
• Tests were performed on a virtualized server with 4 2.6 GHz Intel
CPUs with 8 threads
• Chosen TPTP Problems:
• 51 Problems
• 479 Megabytes on disk
• 3,341,984 Formulae
• Test Setup:
19

28. benchmarking
• Tests were performed on a virtualized server with 4 2.6 GHz Intel
CPUs with 8 threads
• Chosen TPTP Problems:
• 51 Problems
• 479 Megabytes on disk
• 3,341,984 Formulae
• Test Setup:
• The deduction server ran in the single strategy mode using the
same strategy as the plain eprover
19

29. benchmarking
• Tests were performed on a virtualized server with 4 2.6 GHz Intel
CPUs with 8 threads
• Chosen TPTP Problems:
• 51 Problems
• 479 Megabytes on disk
• 3,341,984 Formulae
• Test Setup:
• The deduction server ran in the single strategy mode using the
same strategy as the plain eprover
• Time Limit: 30 Seconds
19

30. benchmarking
• Tests were performed on a virtualized server with 4 2.6 GHz Intel
CPUs with 8 threads
• Chosen TPTP Problems:
• 51 Problems
• 479 Megabytes on disk
• 3,341,984 Formulae
• Test Setup:
• The deduction server ran in the single strategy mode using the
same strategy as the plain eprover
• Time Limit: 30 Seconds
• Memory Limit: 1024 MB
19

31. benchmarking
• Tests were performed on a virtualized server with 4 2.6 GHz Intel
CPUs with 8 threads
• Chosen TPTP Problems:
• 51 Problems
• 479 Megabytes on disk
• 3,341,984 Formulae
• Test Setup:
• The deduction server ran in the single strategy mode using the
same strategy as the plain eprover
• Time Limit: 30 Seconds
• Memory Limit: 1024 MB
• Time was recorded after 5, 10, 20, 30, 40 and all 51 problems
19

32. benchmarking
20

33. deduction as a service
• Having a hosted deduction server will allow users to connect to
the remote server and have their own sessions
21

34. deduction as a service
• Having a hosted deduction server will allow users to connect to
the remote server and have their own sessions
• A win for:
21

35. deduction as a service
• Having a hosted deduction server will allow users to connect to
the remote server and have their own sessions
• A win for:
• Users who cannot install the prover due to compatiblity issues
21

36. deduction as a service
• Having a hosted deduction server will allow users to connect to
the remote server and have their own sessions
• A win for:
• Users who cannot install the prover due to compatiblity issues
• Users who have limited resources
21

37. deduction as a service
• Having a hosted deduction server will allow users to connect to
the remote server and have their own sessions
• A win for:
• Users who cannot install the prover due to compatiblity issues
• Users who have limited resources
• Before The protocol, apps used to interacting with E by invoking
the executable in a subprocess. The interaction protocl will
make it much easier to intgerate apps with E
21

38. future work

39. future work
• Controlling client allowed resource
23

40. future work
• Controlling client allowed resource
• Clustering and load balancing
23

41. future work
• Controlling client allowed resource
• Clustering and load balancing
• Supporting other backends
23

42. future work
• Controlling client allowed resource
• Clustering and load balancing
• Supporting other backends
• Supporting other pruning techniques
23

43. conclusion

44. • Benchmarking results shows that the deduction server mode
outperforms plain eprover in large axiom sets with shared
knowledge bases
25

45. • Benchmarking results shows that the deduction server mode
outperforms plain eprover in large axiom sets with shared
knowledge bases
• Offering deduction as a hosted service will allow users to use
remote theorem provers on powerful servers
25

46. • Benchmarking results shows that the deduction server mode
outperforms plain eprover in large axiom sets with shared
knowledge bases
• Offering deduction as a hosted service will allow users to use
remote theorem provers on powerful servers
• The interaction protocol makes integrating apps with E much
easier
25

47. Questions?
26

48. Thank you
27

49. references I
Opencyc — Wikipedia, the free encyclopedia.
[Online; accessed 28-July-2015].
28