Arnon Boneh's PREDUCE. The tl;dr is you have some polytope in n-dimensional euclidean space that represents a linear optimization problem. The polytope is overdefined (as is wont in linear optimization, you sometimes have several unnecessary constraints on the polytope becaue of the nature of the problem) and so PREDUCE is a way to determine which constraints are redundant and which are necessary in a probablistic manner. The way it works is you generate points at random inside of the euclidean space and determine which constraints it satisfies and fails. Using that information and some computation you can eliminate some of the constraints.

In low dimensional (<5 variables or so) or low complexity (<20 constraints or so) polytopes you don't generally see much benefit from this, but in high dimensional or high complexity polytopes you can sometimes reduce computation time solving the optimization problem significantly.