Bilevel Decomposition Fixes CMA-ES on Mixed Categorical-Continuous Problems

Standard CMA-ES chokes on problems where categorical and continuous decisions interact strongly—its search distribution treats them independently, so joint structure goes unmodeled. The new arXiv paper 2606.12885 from an anonymous team fixes that with a bilevel decomposition that explicitly conditions continuous search on categorical choices.

Two Loops Beat One Loop

The outer loop optimizes over categorical configurations using information-geometric stochastic relaxation. For each candidate categorical setting, an inner loop runs a separate continuous optimization—effectively making the continuous search aware of which categorical bucket it’s in. This mirrors the natural structure of real-world problems where, say, material type dictates different optimal manufacturing parameters.

Running two nested optimizations normally costs double, so the authors introduce a warm-starting strategy: cache the best continuous solutions from previous categorical configurations and seed the inner loop with the best cached candidate. After each outer iteration, the cache updates. That cuts the bilevel overhead enough to make the method computationally cheaper than existing mixed-variable CMA-ES variants on the benchmarks tested.

Binary-Continuous Results That Matter

Experiments on binary-continuous domains—the hardest subset of mixed categorical-continuous problems—show the method outperforms state-of-the-art CMA-ES extensions in handling interactions. The paper includes both previously reported and newly proposed interaction types, so the improvement isn’t cherry-picked. For a practitioner running simulation-based optimization (e.g., hyperparameter tuning with nested scaling choices), this means fewer expensive evaluations to find the real optimum.

The framework is generic: any evolution strategy with an information-geometric foundation could slot into either loop. Expect this to show up in robust hyperparameter optimization libraries and simulation-driven design tools within the year.

Source: Mixed-Categorical Black-Box Optimization via Information-Geometric Bilevel Decomposition
Domain: arxiv.org