Edinburgh and Calculemus

1
Edinburgh and Calculemus

• Simon Colton
• Universities of Edinburgh and York

2
Generating Conjectures About Maple Functions
(Task 2.2)

• Simon Colton
• Universities of Edinburgh and York

3
Some Facts

• Mathematicians dont use ATP or ATF
• But they do use CAS
• CAS help them make discoveries
• The HR program (ATF) makes discoveries
• Interesting discoveries are hard to prove
• ATP only proves easy things
• But easy is the opposite of hard

4
So, what to suggest?

• HR is given maple library functions
• HR makes many conjectures
• Otter proves some of them
• Throw these away too easy
• User chooses conjectures as axioms
• Use Otter to prove more easy theorems

5
Calculemus Paid For

• Getting HR into Mathweb
• Talk to both Otter and Maple there
• Also application to TPTP library
• Work done at Saarbrucken
• Enabled HR to interface with Maple
• Crash course in Maple programming
• Work done at Karlsruhe

6
Example Number Theory

• Maple numtheory package, integers 1-100
• tau, sigma, isprime functions
• HR exhausted up to complexity 6
• Compose, exists and split production rules
• Otter given 10 seconds to prove theorems
• Only ground instances supplied, e.g., isprime(3).
• Are there any interesting relationships?
• If so, can the user find them in HRs output?

7
Results 1

• Takes two minutes to finish
• 180 implicate conjectures produced
• 82 of them proved by Otter
• all a,b (tau(a)b sigma(b)a isprime(b)
• ? exists c (sigma(b)c tau(c)b))
• Otter removes nearly half the conjs.

8
Results 2

• User goes through
• Chooses 10 (true) conjectures
• E.g., tau(a)2 ? isprime(a)
• These are added as axioms
• All conjectures proved again
• Yield reduces to just 29 conjectures
• Possible for user to check through these

9
Results 3

• Whats left?
• isprime(tau(a)) ? isprime(tau(tau(a))
• isprime(sigma(a)) ? isprime(tau(a)
• isprime(tau(a)) ? tau(sigma(tau(a)))tau(a)
• User can order by applic/surprise
• Still some uninteresting ones
• But were working on that

10
Future Work

• Get all this working inside Mathweb
• Make HR more interactive
• Ask user on the fly whether a conj. is true
• Work with Roy McCasland
• EPSRC Visiting Research Fellow
• Apply HR to Zariski Spaces

11
Three Possible Challenge Problems (Task 3.5)

• Simon Colton
• Universities of Edinburgh and York

12
Overview

• Clarkes Analytica problems
• Jurgen Zimmer, Alan Bundy et Al
• Lawveres Conceptual Mathematics
• Alan Smaill
• Sutcliffes TPTP library
• Simon Colton, Jurgen Zimmer, Geoff Sutcliffe

13
Clarkes Analytica Problems

• Analytica
• Theorem prover over Mathematica
• Written by Ed. Clarke
• Challenge problems
• Analysis problems (e.g., continuity)
• Analytica able to prove the
• Jürgen has already talked about this

14
Lawveres Conceptual Maths

• This book, while being an introductory text,
  gives insight into the search for solutions and
  abstraction of ideas that goes on in mathematical
  practice, in a way unusual in mathematical
  texts.  It also gives many interesting abstract
  characterisations of mathematical concepts. Some
  examples Brouwer's fixed point theorems for a
  real interval, and for the closed disc, are
  derivable from a few intuitive properties of
  continuity similarly for Banach's fixed-point
  theorem for contraction maps.  For another
  example, the generality of this approach allows a
  characterisation of the notion of "diagonal
  argument, which is then applicable in many
  versions (e.g., Cantor's diagonal argument, or
  Goedel's diagonalisation in his theorem about
  incompleteness of formal systems.  The challenge
  is to formalise these ideas in such a way as to
  allow exploration of the search spaces
  involved,while retaining the intuitive feel for
  the concepts involved.

15
TPTP Problem Generation

• TPT Problem library
• Roughly 6000 problems
• De facto standard for comparing ATPs
• Maintained by Geoff Sutcliffe
• Problems
• Critics complain that people fine-tune
• ATPs keep getting better
• Vital to keep adding new problems

16
Challenge Problem

• Automatically generate new TPTP probs
• Use the HR program
• Find conjectures empirically
• Computation
• Use ATPs
• To test for theoremhood
• To assess rating (number of ATPs which fail)
• Deduction

17
Results 1

• Approach 1 No interaction
• 46,000 theorems produced by HR in 10 mins
• Sent to Geoff Sutcliffe to assess
• Results
• All provable by SPASS in 120 seconds
• 40 Not provable by E (rating 0.8)
• 144 Not provable by vamp/gand/otter (0.6)
• Challenge beat SPASS

18
Results 2

• HR in MathWeb
• Uses Otter, Spass, E and Bliksem
• 12,000 conjectures tried
• All but 70 were proved
• For each prover
• A theorem which only it could not prove
• Beat SPASS 4 times!

19
Next Stage

• Add more intelligence to choosing conjs
• Add more provers to MathWeb
• Offer HR as a mathematical service
• User supplies axioms
• Requires a certain number of theorems
• Which are of a certain difficult
• For certain provers

20
Edinburghs Proposed Industrial Applications

• Simon Colton
• Universities of Edinburgh and York

21
Application 1 on Platform Grant

• Configuring the Grid
• E-science (large datasets, distributed access)
• Edinburgh is the national centre ()
• Web Description Service Language
• Formally verify that QOS specs are met
• E.g., reliability and redundancy
• Very preliminary

22
Application 2 on Platform Grant

• Bioinformatics
• Biochemical structures as logical expr.
• Composition as inference processes
• Infer compound properties from comp.
• Proof planning to synthesise structures
• Critics to analyse failed attempts
• Again, very preliminary