Stochastic GO algorithms

Abstract

We consider stochastic methods as those algorithms that use (pseudo) random numbers in the generation of new trial points. The algorithms are used a lot in applications. Compared to deterministic methods they are often easy to implement. On the other hand, for many applied algorithms no theoretical background is given that the algorithm is effective and converges to a global optimum. Furthermore, we still do not know very well how fast the algorithms converge. For the effectiveness question, Törn and Žilinskas (1989) already stress that one should sample “everywhere dense”. This concept is as difficult with increasing dimension as doing a simple grid search. In Section 7.2 we describe some observations that have been found by several researchers on the question of increasing dimensions.