Safe Bayesian Optimization for Physical Systems

This direction is merged into another direction: safe exploration. Its last version stays readable and citable here.

Version 1 of 164 papers71 conceptsfrozen 2026-10-04

In traditional machine learning, "failing fast" is a virtue. If an image classifier mislabels a cat as a dog, the cost is negligible. However, in physical domains like robotics, power systems, or medical devices, a single failure can be catastrophic. Tuning a quadrotor's controller by trial and error might lead to a crash; optimizing a high-power laser pulse could shatter expensive optics.

The research direction emerging from these papers focuses on Safe Bayesian Optimization (SafeBO). This framework allows us to optimize unknown, "black-box" functions while guaranteeing that every single sample taken during the process remains within a predefined safety boundary.

The Safety-First Optimization Loop

Standard optimization algorithms, such as those described by Eric Schulz and Andreas Krause in their tutorial on Gaussian processes, typically prioritize the "exploration-exploitation" tradeoff. They explore uncertain regions to find a global optimum. SafeBO adds a third, non-negotiable priority: safety.

The core intuition, pioneered by Yanan Sui and Alkis Gotovos in the SAFEOPT algorithm, is to model the safety constraint itself as a Gaussian Process (GP). Because GPs provide a measure of uncertainty (the "confidence interval"), we can identify a Safe Set: a region of the parameter space where we are statistically certain the system will not violate its limits.

Expand then Maximize

SafeBO works through a two-stage strategy at each iteration. Instead of just looking for the best performance, the algorithm classifies points in the safe set into two categories:

  1. Maximizers: Points that, based on our current model, have the potential to be the global optimum.
  2. Expanders: Points near the boundary of the safe set that, if sampled, are likely to reduce our uncertainty and "expand" the known safe region.

By alternating between expanding the safe set and maximizing the objective within it, algorithms like SAFEOPT-MC—developed by Felix Berkenkamp and Angela P. Schoellig—can safely navigate to the best possible configuration. As noted in Safe controller optimization for quadrotors, this is fundamentally restricted to the reachable set: the collection of points connected to an initial safe seed. If a global optimum is "hidden" behind an unsafe region, the algorithm will correctly refuse to cross that "valley of death" to reach it.

Scaling to the Real World

While the basic SafeBO framework is powerful, real-world systems introduce complexities that require specialized mechanisms:

  • High Dimensions: Bayesian Optimization often struggles as the number of parameters grows. Johannes Kirschner and Mojmír Mutný addressed this with LINEBO, which decomposes high-dimensional problems into a series of safe one-dimensional subspace searches.
  • Digital Twins: To reduce the number of risky physical trials, Mahdi Nobar and John Lygeros propose using Digital Twins. By combining simulation data with real-world corrections, they can guide the optimizer toward safe regions more efficiently.
  • Geometric Guarantees: In automotive control, Johanna Menn and Sebastian Trimpe utilize Lipschitz continuity. By assuming a maximum "rate of change" for the system's response, they can mathematically prove that a small step from a safe point will not result in an unsafe one.

Safety as a Performance Booster

A recurring theme across these papers is that safety constraints are not just a burden—they can actually be a "performance booster." By restricting the search space to feasible regions, algorithms like Conformal Policy Control by Drew Prinster and Anqi Liu prevent the agent from wasting time exploring low-quality or impossible actions.

Whether it is preventing wildfires in power grids (Bistra Dilkina) or ensuring chemical reactions don't become hazardous (Dong Han), the shift toward SafeBO represents a move from "black-box" trial-and-error to a principled, risk-aware exploration of the physical world.

Papers in this version

Showing 60 of 64 papers.

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_safe_bayesian_optimization_for_physical__020e2c21,
  title = {Safe Bayesian Optimization for Physical Systems},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/safe-bayesian-optimization-for-physical-systems/rdr_020e2c21cdc84700b7e65e7ca96dd5fb},
  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_safe_bayesian_optimization_for_physical__020e2c21_v1,
  title = {Safe Bayesian Optimization for Physical Systems},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/safe-bayesian-optimization-for-physical-systems/rdr_020e2c21cdc84700b7e65e7ca96dd5fb/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}