Optimization Across Mixed Continuous and Categorical Search Spaces

You are reading version 1 of this direction, frozen 2026-10-04. The latest version is version 2.

Version 1 of 2448 papers606 conceptsfrozen 2026-10-04

Most optimization algorithms are built on the assumption of a smooth, continuous landscape. If you nudge a parameter slightly to the right, the objective function should change by a predictable amount. This allows us to use gradients, calculate distances, and build surrogate models like Gaussian Processes.

However, real-world engineering and machine learning problems rarely offer such a tidy environment. When designing a new material, you might need to choose between a set of discrete polymer types (categorical) while simultaneously tuning the filler concentration (continuous). When configuring an AutoML pipeline, you must select an algorithm like a Random Forest or an SVM (categorical) and then optimize its specific hyperparameters (continuous).

This "mixed-variable" headache creates a fundamental mismatch: how do you calculate the "distance" between a Support Vector Machine and a Decision Tree? How do you take the gradient of a hairstyle choice in a character creator? Researchers are bridging this gap by developing frameworks that treat discrete and continuous variables as a unified search space.

Mapping Categories to Latent Spaces

One of the most intuitive ways to handle qualitative choices is to stop treating them as labels and start treating them as coordinates. Wei Chen and Daniel W. Apley demonstrated this in Bayesian Optimization for Materials Design with Mixed Quantitative and Qualitative Variables.

Instead of forcing a Gaussian Process to understand a category like "Polymer A," they map these choices into a continuous latent space. In this space, materials that behave similarly are positioned close together. This allows the optimizer to use standard continuous kernels to "smooth" the categorical gaps. Akshay Iyer and colleagues applied this data-centric approach to nanocomposite design, showing that these latent mappings not only speed up discovery but also reveal physical intuitions about how different materials relate to one another.

Decomposing the Problem with ADMM

If a problem is too messy to solve all at once, the next best strategy is to break it into pieces that we do know how to solve. Parikshit Ram and Sijia Liu pioneered this using the Alternating Direction Method of Multipliers (ADMM).

In their framework for AutoML pipeline configuration, they decompose the search into three distinct sub-problems:

  1. Algorithm Selection: A discrete choice.
  2. Hyperparameter Optimization: A continuous search.
  3. Integer Projection: Handling variables that must be whole numbers.

By alternating between these steps, the algorithm can use specialized tools for each variable type—like gradient descent for numbers and combinatorial search for categories—while ensuring the global solution remains coordinated. This approach is particularly powerful because it can handle "black-box" constraints like latency or fairness that are often ignored by simpler tuners.

Another strategy is to use different "engines" for different parts of the search. Patrick Koch and the team behind Autotune combine Genetic Algorithms (GA) with Generating Set Search.

The intuition here is that GAs are excellent at jumping across the "cliffs" of a categorical space—switching from one model type to another—while local refinement methods can then "polish" the continuous parameters once a promising region is found. This hybrid approach avoids the common trap where an optimizer gets stuck in a sub-optimal categorical choice because it cannot easily "jump" to a better one.

Bridging Structure and Parameters

A particularly difficult version of the mixed-variable problem appears in Symbolic Regression and Scientific Discovery. Here, the "discrete" part isn't just a category; it's a mathematical structure or a molecular graph.

In Deep Symbolic Optimization, the framework uses reinforcement learning to "write" the discrete structure of an equation. However, an equation like $y = \sin(cx)$ is useless without the right constant $c$. Pingchuan Ma addresses this in the Scientific Generative Agent (SGA) by using a bilevel optimization: a Large Language Model proposes the discrete symbolic structure, while a differentiable simulation fine-tunes the continuous physical constants.

Unifying Variables as Bandit Arms

Finally, some researchers treat the categorical choice as a high-level decision and the continuous variables as the "payoff" of that decision. Dang Nguyen and Santu Rana developed Bandit-BO, which treats each category as an "arm" in a multi-armed bandit problem.

The algorithm uses Thompson Sampling to decide which category to explore next, and then uses Bayesian Optimization to navigate the continuous parameters within that category. This is especially useful for problems like AutoML, where different algorithms have entirely different sets of hyperparameters (e.g., a "kernel" parameter exists for SVMs but makes no sense for Random Forests).

By moving away from "rounding" discrete variables or treating categories as independent silos, these mixed-variable frameworks allow us to optimize the complex, heterogeneous systems that define modern science and engineering.

Papers in this version

Showing 60 of 448 papers.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 193 papers
  2. Version 1 · 2026-10-04 · 448 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_mixed_variable_optimization_39c4ebeb,
  title = {mixed-variable optimization},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/mixed-variable-optimization/rdr_39c4ebeb27e34306901507337b3b7f15},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_optimization_across_mixed_continuous_and_39c4ebeb_v1,
  title = {Optimization Across Mixed Continuous and Categorical Search Spaces},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/optimization-across-mixed-continuous-and-categorical-search-spaces/rdr_39c4ebeb27e34306901507337b3b7f15/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}