Abstract

Optimization problems, which seek to minimize or maximize some value, are ubiquitous in large scale computing. They are also among the most computationally difficult problems to solve. Consequently, many practical applications rely on approximations, settling for "good enough" when the "absolutely best" answer cannot be feasibly attained. Genetic algorithms, which simulate the evolution of potential solutions toward better and better alternatives, are a very powerful stochastic technique for finding approximate solutions to optimization problems. However, their limitations are currently not well understood. This leaves the programmer little guidance as to when this versatile technique should be used and when it should not. Recently, surprising theoretical results have led to a classification of optimization problems according to how well they can be approximated by reasonably fast algorithms. These limitations are inherent in the problem, and are therefore implementation independent. T...

Citations