Automated Discovery in Pure Mathematics

Download Report

Transcript Automated Discovery in Pure Mathematics

Automated Discovery in
Pure Mathematics
Simon Colton
Universities of Edinburgh and York
Overview of Talk
Some example discoveries

ATP, CSP, CAS, ad-hoc methods
The HR system


Automated theory formation
Overview of applications
Application to mathematical discovery

Finite algebras, number theory, refactorables
Demonstration

NumbersWithNames program
Automated Discoveries #1
Robbins algebras are boolean

Automated theorem proving, McCune+Wos
Quasigroup existence problems (QG6.17)

Constraint solvers, John Slaney et al.
Inconsistency in Newton’s Principia

Formal methods (NS-analysis), Fleuriot
Automated Discoveries #2
Mersenne prime: 26972593 – 1

Distributed (internet) search, CAS
New geometry results

Chou using Wu’s method
Simple axiomatisations of algebras


Group: x(y(((zz-1)(uy)-1)x))-1=u
McCune and Kunen, ATP
Automated Discoveries #3
Fajtlowicz’s Graffiti graph theory program


All G, Chrom+Rad < MaxDeg+FreqMaxDeg
60+ papers about it’s conjectures
Bailey’s PSQL algorithm
New formula for :
i (1/16i)(4/(8i+1)-2/(8i+4)-1/(8i+5)-1/(8i+6))
 Easier to calculate nth hex digit of 

Theories in Pure Mathematics
Concepts

Examples and definitions
Statements

Conjectures and theorems
Explanations

Proofs, counterexamples
e.g., pure maths:group theory




Concepts: cyclic groups, Abelian groups
Conjecture: cyclic groups are Abelian
Examples provide empirical evidence
Simple proof for explanation
HR: Theory Formation Cycle
Start with background knowledge

1.
2.
3.
4.
5.
user-supplied axioms + concepts
Invent a new concept (machine learning)
Look for conjectures empirically (d-mining)
Prove the conjectures (theorem proving)
Disprove the conjectures (model generation)
Assess all concepts w.r.t. new concept
1. Invent a new concept

Build it from the most interesting old concepts
Inventing New Concepts
Ten General Production Rules (PR)


Work in all domains (math + non math)
Build new concept from one (or two) old ones
Example: Abelian groups




Given: [G,a,b,c] : a*b=c
Compose PR: [G,a,b,c] : a*b=c & b*a=c
Exists PR: [G,a,b] :  c (a*b=c & b*a=c)
Forall PR: [G] :  a b ( c (a*b=c & b*a=c))
Making Conjectures
Theory formation step

Attempt to invent a new concept
Concept has same examples as previous one

HR makes an equivalence conjecture
Concept has no examples

HR makes a non-existence conjecture
Examples of one concept are all examples of
another concept

HR makes an implication conjecture
Proving Theorems
HR relies on third party theorem provers
Equivalence conjectures:




Sets of implication conjectures
From which prime implicates are extracted
E.g.  a (a*a=a  a=id)
a*a=a  a=id,
a=id  a*a=a
HR uses the Otter theorem prover


William McCune et al.
Only uses this for finite algebras
Disproving Non-Theorems
Any conjectures which Otter can’t prove



HR looks for a counterexample
Using the MACE model generator
Also written by William McCune
Other possibilities:

Computer algebra, constraint satisfaction
Counterexamples are added to the theory

Fewer similar non-theorems are made later
Assessing Interestingness
New concepts from interesting old ones
Concepts measured in terms of:


Intrinsic values, e.g. complexity of definition
Relational values, e.g. novelty of categorisation
Concepts also assessed by conjectures

Quality, quantity of conjectures involving conc.
Conjectures also assessed


Difficulty of proof (proof length from Otter)
Surprisingness (of LHS and RHS definitions)
Bootstrapping ATF Cycle
Applications of HR
Puzzle generation

Next in sequence, odd one out
Automated theorem proving

Discovering useful lemmas
Constraint satisfaction problems

Discovering additional constraints
Machine learning tasks

Puzzle solving, prediction tasks
Studying machine creativity

Multi-agent, cross-domain, meta-level
Application to
Mathematical Discovery
Exploration of algebras using HR


Anti-associative algebras
Quasigroups
Number theory results


Encyclopedia of Integer Sequences
Using HR and NumbersWithNames
Refactorable numbers

Results and open conjectures
Problem solving (Zeitz numbers)
Anti-associative Algebras
(Novel domain to me)
all a,b,c a*(b*c)  (a*b)*c
Used HR with Otter and MACE (2 hours)
34 examples, sizes 2 to 6 (exists each size)
AAAs are not: abelian or quasigroups

Quasigroups must have associative triple
Have two elements on diagonal
Have no identity, or even local identity
Commutative pairs are not co-squares
Quasigroup Results
Part of CSP project
QG3 quasigroups: (a*b)*(b*a)=a
HR conjectured, Otter proved, We interpreted



Diagonal elements are all different
a*a=b  b*b=a
a*b=b  b*a=a
QG3 quasigroups are anti-Abelian


a*b = b*a  a=b
Corollary to one of HR’s results (with our help)
10x speed up over naïve model
Neil Sloane’s Encyclopedia
of Integer Sequences
Large database of sequences

E.g., Primes: 2, 3, 5, 7, 11, 13,…
Contains 67,000+ sequences (36 years)

A new sequence must be novel, infinite, interesting

HR has invented 20 new sequences



All supplied with interesting theorems (our proof)
Datamining the Encyclopedia itself
NumbersWithNames program (details ommitted)
Some Nice Results
Number of divisors, (n), is a prime


2, 3, 4, 5, 7, 9, 11, 13, …
m(n) is prime  (n) is prime
g(n) = #squares dividing n

1, 1, 1, 2, 1, 1, 1, 2, 2, 1, 1, 2, …
numbers setting the record for g(n)


1, 4, 16, 36, 144, 576, …
Squares of the highly composite numbers
Perfect numbers are pernicious
Refactorable Numbers
Number of divisors is itself a divisor



1, 2, 8, 9, 12, 18, 24, 36, 40, …
HR’s first success [not in Encyclopedia]
Turned out to be a re-invention (1990)
Preliminary results (* - made by HR)





Infinitely many refactorables
Odd refactorables are perfect squares *
Congruent to 0, 1, 2 or 4 mod 8 *
Perfect numbers are not refactorable *
m,n relprim and refactorable  mn
refactorable
Refactorables – Deeper Results
Natural density is zero

Kennedy and Cooper 1990
Joshua Zelinsky (hot off the press)



T(n) < 0.5 B(n) with finitely many
counterexamples (max 1013)
T(n) = #refacs < n, B(n) = #primes < n
Assuming Goldbach’s strong conjecture
 Every integer is the sum of 5 or fewer refactorables
Zelinsky uses the results from HR
Refactorables – Questions…..
Numbers n!/3 are refactorable*
Numbers for which ((n))=n are refactorable*
(x) = #integers less than or equal to and coprime to x
There are infinitely many pairs of refactorables

(1,2), (8,9), (1520,1521), (50624,50625), …
There are no triples of refactorables


We know there are no quadruples
And no triples less than 1053
Demonstration – Zeitz numbers
Hungarian maths competition
Multiply four consecutive numbers


n(n+1)(n+2)(n+3)
Never a square number
Demonstration

Using NumbersWithNames
Future Work: HR Project
McCasland?

Use HR to explore Zariski spaces
Colton: Express HR as a ML program

Try domains other than maths (bioinformatics)
Walsh: Integrate HR


With every maths program ever written
In particular Maple computer algebra
Bundy:

Build an automated mathematician
Web Pages
HR:

www.dai.ed.ac.uk/~simonco/research/hr
NumbersWithNames program:

www.machine-creativity.com/programs/nwn
Encyclopedia of Integer Sequences:

www.research.att.com/~njas/sequences