REMEDA: Random Embedding EDA for Optimising Functions with Intrinsic Dimension

Abstract

It has been observed that in many real-world large scale problems only few variables have a major impact on the function value: While there are many inputs to the function, there are just few degrees of freedom. We refer to such functions as having a low intrinsic dimension. In this paper we devise an Estimation of Distribution Algorithm (EDA) for continuous optimisation that exploits intrinsic dimension without knowing the influential subspace of the input space, or its dimension, by employing the idea of random embedding. While the idea is applicable to any optimiser, EDA is known to be remarkably successful in low dimensional problems but prone to the curse of dimensionality in larger problems because its model building step requires large population sizes. Our method, Random Embedding in Estimation of Distribution Algorithm (REMEDA) remedies this weakness and is able to optimise very large dimensional problems as long as their intrinsic dimension is low.