Research News recently highlighted self-improving software developed by
Princeton computer science Professor Bernard Chazelle and his students.
people are expected to learn on the job, why isn't software?” writes
TRN. “Although some kinds of software are capable of learning, it's
more difficult to design software that learns as it works without
requiring a separate training process.
Bernard Chazelle, professor in the
Computer Science department
“Princeton University researchers have designed algorithms -- the
logic underlying software -- that learn from data that they don't know
anything about ahead of time and then tune themselves to better handle
those types of data. The key is that the algorithms learn from how the
pieces of data fit within the range of possibilities, rather than
having to learn the data's details.
“It turns out that even though any given piece of
data is random, individual pieces fall into relatively narrow ranges
that an algorithm can learn from. An algorithm can also improve after
learning from a relatively small number of samples.
researchers built two self-improving algorithms, a sorting algorithm
and a clustering algorithm. Sorting algorithms put pieces of data into
some type of order and clustering algorithms group like pieces of data.
“The algorithms promise to be forerunners of software that
alters its default configuration on its own as it learns how it is
Chazelle and graduate students Nir Ailon, Ding Liu, and Seshadhri Comandur wrote a paper (pdf) about their work, which Comandur presented at an ACM-SIAM Symposium on Discrete Algorithms in January.
For Princeton Engineering media relations, EQuad News, news releases:
Chazelle and graduate students Nir Ailon, Ding Liu, and Seshadhri Comandur wrote a paper (pdf) about their work, presented at an ACM-SIAM Symposium on Discrete Algorithms in January.