CatMADS: Mesh Adaptive Direct Search for constrained blackbox optimization with categorical variables

TL;DR

CatMADS extends the MADS optimization algorithm to effectively handle mixed-variable problems with categorical, integer, and continuous variables, providing strong theoretical guarantees and demonstrating superior empirical performance.

Contribution

It introduces a generalized MADS framework for mixed-variable constrained optimization, incorporating categorical variables and convergence analysis.

Findings

CatMADS outperforms state-of-the-art solvers on 32 mixed-variable problems.

The framework ensures Clarke stationarity and handles constraints effectively.

Empirical results confirm the efficiency and robustness of CatMADS.

Abstract

Solving optimization problems in which functions are blackboxes and variables involve different types poses significant theoretical and algorithmic challenges. Nevertheless, such settings frequently occur in simulation-based engineering design and machine learning. This paper extends the Mesh Adaptive Direct Search (MADS) algorithm to address mixed-variable problems with categorical, integer and continuous variables. MADS is a robust derivative-free optimization framework with a well-established convergence analysis for constrained quantitative problems. CatMADS generalizes MADS by incorporating categorical variables through distance-induced neighborhoods. A detailed convergence analysis of CatMADS is provided, with flexible choices balancing computational cost and local optimality strength. Four types of mixed-variable local minima are introduced, corresponding to progressively stronger notions of local optimality. CatMADS integrates the progressive barrier strategy for handling constraints, and ensures Clarke stationarity. An instance of \catmads employs cross-validation to construct problem-specific categorical distances. This instance is compared to state-of-the-art solvers on 32 mixed-variable problems, half of which are constrained. Data profiles show that CatMADS achieves the best results, demonstrating that the framework is empirically efficient in addition to having strong theoretical foundations.