Safe Bayesian Optimization Across Noise Models via Scenario Programming

Abdullah Tokmak, Thomas B. Schön, Dominik Baumann

TL;DR

This paper introduces a flexible framework for Safe Bayesian Optimization (BO) that works across diverse noise models, including "difficult" ones like heavy-tailed or parameter-dependent (heteroscedastic) noise. By using scenario programming, the authors replace rigid Gaussian assumptions with a data-driven approach that provides rigorous safety guarantees and convergence proofs for real-world, safety-critical systems.

What Problem Does This Paper Address?

Bayesian Optimization is a go-to method for tuning controllers in robotics because it is sample-efficient. However, in safety-critical applications (like a robot arm moving near humans), we need "Safe BO," which ensures the system never violates constraints during the tuning process.

The "Achilles' heel" of most existing Safe BO algorithms (like the popular SAFEOPT) is their reliance on the assumption of homoscedastic sub-Gaussian noise. In plain English: they assume the measurement noise follows a nice, predictable bell curve that doesn't change regardless of what the robot is doing.

In reality, sensors are messy. Network delays, LiDAR interference, and hardware vibrations often produce "heavy-tailed" noise (outliers are more common than a Gaussian predicts) or "heteroscedastic" noise (the noise level changes depending on the input parameters). If a "safe" algorithm ignores these realities, it will be too optimistic about its safety margins, leading to catastrophic real-world failures.

Methods

The authors propose a new algorithm, Safe BO under general noise, which bridges the gap between theoretical safety and messy real-world noise.

  1. Scenario Programming: Instead of assuming a specific mathematical shape for the noise, the authors use the "scenario approach." If you can sample the noise (either from a model or historical data), you can solve a simple optimization problem to find a high-probability bound on that noise.
  2. Simultaneous Bounds: They provide a way to adjust these bounds over time (Algorithm 1) so that the safety guarantee holds across all iterations of the optimization, not just a single step.
  3. Modified Confidence Intervals: They integrate these scenario-based bounds into the Gaussian Process (GP) surrogates used in BO. This allows the algorithm to build "safe sets" that are robust to the specific noise profile of the system.
  4. Safety & Optimality Proofs: The paper mathematically proves that the algorithm will remain safe with high probability and eventually find the optimal parameters (within a specified tolerance $\delta$).

Datasets

The paper evaluates the algorithm using:

  • Synthetic Benchmarks: Functions sampled from a Reproducing Kernel Hilbert Space (RKHS) with simulated heteroscedastic heavy-tailed noise (specifically, Student’s-t distributions scaled by the input).
  • Franka Emika Manipulator Simulation: A high-fidelity simulation using the MuJoCo physics engine. The goal was to tune an operational space impedance controller (LQR-based) for a set-point tracking task while ensuring the robot doesn't exceed distance-to-target constraints.

Key Insights and Ideas

  • Sampling over Formulas: The core "aha moment" is that you don't need a closed-form analytical expression for your noise to be rigorous. If you can sample the noise distribution, scenario theory gives you the mathematical "teeth" to guarantee safety.
  • Robustness vs. Optimism: Standard Safe BO algorithms are "optimistic"—they assume the noise is well-behaved. This paper shows that when noise is heavy-tailed, this optimism leads to safety violations. Their approach is "cautious" in exactly the right way.
  • Unified Framework: This isn't just a "heavy-tailed noise" paper; it’s a framework. The same algorithm can be toggled between standard Gaussian noise and complex, input-dependent noise models just by changing how the "scenarios" (noise samples) are generated.

Key Results

  • Safety Preservation: In synthetic tests with heavy-tailed noise, classic Safe BO algorithms frequently violated safety constraints (indicated by red crosses in the paper's figures). The proposed algorithm maintained safety throughout the entire process.
  • Franka Emika Robot Tuning: The algorithm successfully tuned the robot's controller in 200 iterations. It started from a safe but low-performance initial guess and converged to a high-reward parameter set (a "white star" in their plots) without a single safety violation during the learning phase.
  • Computational Efficiency: The number of noise samples ($m_t$) required for the scenario approach scales linearly with the desired accuracy ($\nu$) but only logarithmically with the confidence level ($\kappa$). This means you can get extremely high confidence (e.g., 99.9%) without a massive computational overhead.

Limitations and Future Work

  • The Sampling Requirement: The algorithm assumes you can sample from the noise distribution (Assumption 1). While this is easier than deriving a formula, it still requires either a good simulator or enough prior data to build a "noise oracle."
  • Exploration Speed: Because the algorithm is more cautious to ensure safety under "weird" noise, it may explore the parameter space more slowly than algorithms that assume simple Gaussian noise.
  • Future Directions: The authors suggest moving toward data-driven noise oracles—using machine learning to learn the noise model on the fly—rather than relying on pre-defined analytical noise models. They also aim to investigate the growth rate of the confidence parameters more deeply to further optimize convergence speed.

Summary updated

Abstract

Safe Bayesian optimization (BO) with Gaussian processes is an effective tool for tuning control policies in safety-critical real-world systems, specifically due to its sample efficiency and safety guarantees. However, most safe BO algorithms assume homoscedastic sub-Gaussian measurement noise, an assumption that does not hold in many relevant applications. In this letter, we propose a straightforward yet rigorous approach for safe BO across noise models, including homoscedastic sub-Gaussian and heteroscedastic heavy-tailed distributions. We provide a high-probability bound on the measurement noise via the scenario approach, integrate these bounds into high probability confidence intervals, and prove safety and optimality for our proposed safe BO algorithm. We deploy our algorithm in synthetic examples and in tuning a controller for the Franka Emika manipulator in simulation.
BibTeX
@inproceedings{tokmak2025safe,
  author = {Abdullah Tokmak and Thomas B. Schön and Dominik Baumann},
  title = {{Safe Bayesian Optimization Across Noise Models via Scenario Programming}},
  booktitle = {IEEE Control. Syst. Lett},
  year = {2025},
  url = {https://openreview.net/forum?id=RLSoCZpVOW},
}

Loading related content...

Figure