Transcript Slide 1
2
G ENETIC A LGORITHMS FOR
FAST M ATRIX M ULTIPLICATION
András Joó
Anikó Ekárt
Juan Neirotti
United Kingdom
GECCO 2011 HUMIES AWARDS
14/07/2011
T HE P ROBLEM : R ECURSIVE
M ATRIX M ULTIPLICATION
3
Standard algorithm for multiplying two square
matrices of size n n requires n3 multiplications and
2
dn n 1 additions
Strassen’s algorithm reduces the number of required
multiplications to n log 7 if n is a power of 2
(1969)
2
GECCO 2011 HUMIES AWARDS
14/07/2011
K NOWN L IMITS
4
For matrices of size 22 at least 7 multiplications
needed
For matrices of size 3 3 at least 19 multiplications
needed
Best known exact algorithm for size 3 3 contains
23 multiplications
GECCO 2011 HUMIES AWARDS
14/07/2011
P RACTICAL S IGNIFICANCE
5
An exact algorithm using 22 multiplications on
matrices of size 3 3 would be an improvement on the
best known algorithm for this size
An exact algorithm using 21 multiplications on
matrices of size 3 3 would be an overall improvement
on how recursive matrix multiplication is currently
performed on large matrices
As the search space has size 2.25e+180 for 21
multiplications and 8.71e+188 for 22 multiplications,
respectively, it is highly unlikely that a human or a
simple algorithm would discover a solution!
GECCO 2011 HUMIES AWARDS
14/07/2011
O UR SOLUTION :
PARALLEL GA
6
Parallel island model, with unidirectional ring
topology and migration
Steady-state elitist GA
Continuous real-valued representation
Variety of crossover and mutation operators
Periodic explicit enforcing of diversity
GECCO 2011 HUMIES AWARDS
14/07/2011
GA R ESULTS
7
On matrices of size 3 3
reproduced a solution with 23 multiplications
found an approximate solution of fitness 0.9978
for 22 multiplications
GECCO 2011 HUMIES AWARDS
14/07/2011
W HY H UMAN -C OMPETITIVE ?
8
In 1976, J. D. Laderman published his article
“A noncommutative algorithm for multiplying 3 3
matrices using 23 multiplications” in the Bulletin of the
American Mathematical Society . Others published
equivalent algorithms.
The theoretically proven lower bound is
19 multiplications, but no exact algorithm with less than
23 multiplications is known to date.
Our GA approach could reproduce matrix multiplication
algorithms using 23 multiplications and also led to an
approximate algorithm requiring 22 multiplications.
GECCO 2011 HUMIES AWARDS
14/07/2011
W HICH C RITERIA ?
9
B: The result is equal to or better than a result that was
accepted as a new scientific result at the time when it was
published in a peer-reviewed scientific journal.
D: The result is publishable in its own right as a new
scientific result independent of the fact that the result was
mechanically created.
F: The result is equal to or better than a result that was
considered an achievement in its field at the time it was
first discovered.
G: The result solves a problem of indisputable difficulty in
its field.
GECCO 2011 HUMIES AWARDS
14/07/2011