Learning functions of k relevant variables

作者:

Highlights:

摘要

We consider a fundamental problem in computational learning theory: learning an arbitrary Boolean function that depends on an unknown set of k out of n Boolean variables. We give an algorithm for learning such functions from uniform random examples that runs in time roughly (nk)ωω+1, where ω<2.376 is the matrix multiplication exponent. We thus obtain the first-polynomial factor improvement on the naive nk time bound which can be achieved via exhaustive search. Our algorithm and analysis exploit new structural properties of Boolean functions.

论文关键词:Computational learning theory,PAC learning,Boolean functions

论文评审过程:Received 22 August 2003, Revised 6 April 2004, Available online 4 June 2004.

论文官网地址:https://doi.org/10.1016/j.jcss.2004.04.002