A simple parallel surrogate optimization method over mixed design spaces/dsopt

Abstract

dsopt is an open-source Python package implementing a practical, parallel surrogate-optimization workflow for expensive black-box functions over mixed-type design spaces. It treats continuous, integer and categorical variables uniformly by representing each parameter as a finite set of admissible values. dsopt combines fast XGBoost surrogates, lightweight uncertainty metrics, a Monte-Carlo candidate pool and a sampling strategy that mixes exploitative, Pareto and explorative points to form parallel evaluation batches without costly inner acquisition solvers. With predictable runtimes and minimal dependencies, dsopt makes surrogate-assisted optimization accessible to researchers and practitioners with basic Python skills.