Designing an Optimal Search Algorithm with Respect to Prior Information

Abstract

AbstractThere are many optimization algorithms, most of them with many parameters. When you know which family of problems you face, you would like to design the optimization algorithm which is the best for this family (e.g., on average against a given distribution of probability on this family of optimization algorithms). This chapter is devoted to this framework: we assume that we know a probability distribution, from which the fitness function is drawn, and we look for the optimal optimization algorithm. This can be based (i) on experimentations, i.e. tuning the parameters on a set of problems, (ii) on mathematical approaches automatically building an optimization algorithm from a probability distribution on fitness functions (reinforcement learning approaches), or (iii) some simplified versions of the latter, with more reasonable computational cost (Gaussian processes for optimization).KeywordsFitness FunctionGaussian ProcessMarkov Decision ProcessSampling CriterionLinear ConvergenceThese keywords were added by machine and not by the authors. This process is experimental and the keywords may be updated as the learning algorithm improves.