Studies on stochastic optimisation and applications to the real world

Vincent Berthier 1
1 TAU - TAckling the Underspecified
LRI - Laboratoire de Recherche en Informatique, UP11 - Université Paris-Sud - Paris 11, Inria Saclay - Ile de France, CNRS - Centre National de la Recherche Scientifique : UMR8623
Abstract : A lot of research is being done on Stochastic Optimisation in general and Genetic Algorithms in particular. Most of the new developments are then tested on well know testbeds like BBOB, CEC, etc. conceived to exhibit as many pitfalls as possible such as non-separability, multi-modality, valleys with an almost null gradient and so on. Most studies done on such testbeds are pretty straightforward, optimising a given number of variables for the recognized criterion on the testbed. The first contribution made here is to study the impact of some changes in those assumptions, namely the effect of supernumerary variables that don’t change anything to a function evaluation on the one hand, and the effect of a change of the studied criterion on the other hand. A second contribution is in the modification of the mutation design for the algorithm CMA-ES, where we will use Quasi-Random mutations instead of purely random ones. This will almost always result in a very clear improvement of the observed results. This research also introduces the Sieves Method well known in statistics, to stochastic optimisers: by first optimising a small subset of the variables and gradually increasing the number of variables during the optimization process, we observe on some problems a very clear improvement. While artificial testbeds are of course really useful, they can only be the first step: in almost every case, the testbeds are a collection of purely mathematical functions, from the simplest one like the sphere, to some really complex functions. The goal of the design of new optimisers or the improvement of an existing one is however, in fine, to answer some real world question. It can be the design of a more efficient engine, finding the correct parameters of a physical model or even to organize data in clusters. Stochastic optimisers are used on those problems, in research or industry, but in most instances, an *optimiser is chosen almost arbitrarily. We know how optimisers compare on artificial functions, but almost nothing is known about their performances on real world problems. One of the main aspect of the research exposed here will be to compare some of the most used optimisers in the literature on problems inspired or directly coming from the real-world. On those problems, we will additionally test the efficiency of quasi-random mutations in CMA-ES and the Sieves-Method.
Complete list of metadatas

Cited literature [125 references]  Display  Hide  Download

https://hal.inria.fr/tel-01668371
Contributor : Vincent Berthier <>
Submitted on : Wednesday, December 20, 2017 - 5:55:47 AM
Last modification on : Tuesday, January 8, 2019 - 8:36:01 AM

File

manuscrit_final.pdf
Files produced by the author(s)

Identifiers

  • HAL Id : tel-01668371, version 1

Citation

Vincent Berthier. Studies on stochastic optimisation and applications to the real world. Numerical Analysis [cs.NA]. Université Paris 11, 2017. English. ⟨tel-01668371⟩

Share

Metrics

Record views

290

Files downloads

133