When short runs beat long runs

What will yield the best results: doing one run n generations long or doing m runs n/m generations long each? This paper presents a technique-independent analysis which answers this question, and has direct applicability to scheduling and restart theory in evolutionary computation and other stochastic methods. The paper then applies this technique to three problem domains in genetic programming. It discovers that in two of these domains there is a maximal number of generations beyond which it is irrational to plan a run; instead it makes more sense to do multiple shorter runs.