Bayesian Optimization (BO) is the gold standard for optimizing expensive, black-box functions—whether you are tuning hyperparameters for a deep neural network, designing a hardware accelerator, or finding the optimal gait for a walking robot.
Traditionally, BO operates tabula rasa. It begins with a uniform or random sampling phase, completely oblivious to any domain knowledge. Yet, in practice, human experts rarely start from scratch. An engineer usually has a strong "gut feeling" or historical data suggesting that a learning rate should be around $10^{-3}$, or that a robot's stable walking parameters lie within a specific narrow window.
The challenge is integration: how do we mathematically inject an expert's subjective prior belief into the rigorous, data-driven framework of Bayesian Optimization? Crucially, how do we do this without permanently derailing the optimization if the expert's intuition happens to be completely wrong?
A powerful class of methods solves this by directly embedding prior probability distributions into different components of the Bayesian optimization loop.
The Anatomy of Prior Injection
To understand how these methods work, we must look at the two core components of Bayesian Optimization: the surrogate model (which estimates the objective function's landscape) and the acquisition function (which decides where to sample next based on that estimate). Researchers have developed distinct strategies to inject priors into both components.
[ Expert Prior π(x) ]
│
┌─────────────────┼─────────────────┐
▼ ▼ ▼
┌─────────────────┐┌───────────────┐┌──────────────────┐
│ Surrogate ││ Acquisition ││ Search Space │
│ Model (GP) ││ Function ││ Geometry │
└─────────────────┘└───────────────┘└──────────────────┘
(Shift Mean/Var) (Multiply Bias) (Warp Dimensions)
1. Reshaping the Acquisition Function
The most direct way to guide an optimizer is to alter its decision-making criteria. Standard acquisition functions, such as Expected Improvement (EI), calculate a utility score based purely on the surrogate model's current uncertainty and predicted value.
In $\pi$BO, Hvarfner et al. et al. propose a simple yet elegant modification: multiplying the standard acquisition function by a user-specified prior distribution $\pi(x)$.
$$\alpha_{\text{modified}}(x) = \alpha_{\text{standard}}(x) \cdot \pi(x)$$
This acts as a soft probabilistic filter. If the surrogate model identifies a region as highly promising, but the expert prior deems it highly unlikely, the combined utility score drops. Conversely, regions favored by both the model and the prior are prioritized for immediate evaluation. As demonstrated by the authors in complex robotic tasks, this simple multiplication steers early exploration toward high-probability zones, drastically reducing the initial evaluation budget.
2. Shifting the Surrogate Model Landscape
Instead of modifying how the optimizer chooses points, another approach modifies what the optimizer believes about the landscape before any data arrives. This is typically achieved by altering the Gaussian Process (GP) surrogate model.
Standard GPs typically use a constant prior mean function (often set to zero). However, Qu et al. et al. introduce a framework called DynMeanBO, which maps an expert prior distribution directly onto the GP's mean function. This forces the surrogate model to naturally expect better performance in regions favored by the expert.
Similarly, in scientific discovery workflows, Yuan et al. et al. leverage the latent domain knowledge of Large Language Models (LLMs) to shift the GP mean function toward preferred regions. Rather than relying on rigid mathematical formulas, the surrogate model's baseline expectation is nudged by the qualitative, text-derived preferences of an AI agent.
Alternatively, the prior can modify the model's covariance (uncertainty) structure. For instance, Smirnova et al. et al. incorporate historical knowledge from related tasks by adding a reliability-based penalty term to the diagonal of the GP covariance matrix. This artificially shrinks the model's uncertainty in regions known to be successful in past datasets, forcing the optimizer to exploit those areas early on.
3. Warping the Geometry of the Search Space
A third, highly intuitive approach is to change the geometry of the search space itself. If an expert believes the optimum lies within a specific narrow band, why treat the entire search space uniformly?
Ramachandran et al. et al. propose a technique that warps the input space based on the expert prior. The transformation mathematically "stretches" high-probability regions and "shrinks" low-probability ones.
Uniform Space: [ Low Prob | High Prob | Low Prob ]
▼ ▼ ▼
Warped Space: [ L.P. ] [ High Probability ] [ L.P. ]
To a standard stationary GP kernel, the stretched region now appears physically larger and more detailed, forcing the optimizer to allocate more sample points there. Unpromising regions are compressed into tiny volumes, causing the optimizer to skip over them quickly.
The Safety Valve: Overriding Bad Advice
The primary risk of introducing human intuition into an automated search is the "flawed expert" problem. If the user's prior is completely wrong, a naive implementation might permanently lock the optimizer out of the true global optimum.
To prevent this, modern prior-injection frameworks incorporate a mathematical "safety valve": dynamic decay.
As explored by Fehring et al. et al., the influence of the prior distribution is engineered to decay over time or as more data points are collected. In acquisition-weighting schemes, the prior distribution $\pi(x)$ is typically raised to a power $\beta_t$ that approaches zero as the iteration count $t$ increases:
$$\alpha(x) \cdot [\pi(x)]^{\beta_t}, \quad \lim_{t \to \infty} \beta_t = 0$$
When $\beta_t \to 0$, the term $[\pi(x)]^{\beta_t}$ becomes a uniform multiplier across the entire space, effectively erasing the prior's bias.
In surrogate-modifying methods like those designed by Qu et al., the shifted mean function dynamically decays toward zero as data accumulates. Because Gaussian Process variance naturally shrinks in regions where real data points are collected, the empirical data eventually overwhelms both the prior mean and the prior variance.
This ensures a elegant transition: the algorithm exploits expert intuition for early speedups, but gracefully reverts to pure, data-driven global exploration if the real-world observations contradict the expert's expectations.
Broad Applications
The philosophy of grounding search heuristics in formal Bayesian priors extends far beyond standard hyperparameter tuning.
- Robotics: Tuning physical systems is expensive and subjects hardware to wear and tear. Early work by Lizotte et al. et al. proved that incorporating domain knowledge via priors accelerates robot gait optimization up to five times. More recently, Okada et al. et al. demonstrated that parameters extracted from human demonstrations can serve as priors to rapidly teach robots compliant stiffness control for contact-rich tasks like opening doors.
- Hardware Design Space Exploration: When designing complex computer systems, Nardi et al. et al. showed that using Beta distributions to model expert intuition during the initial sampling phase drastically reduces the evaluation budget required to find optimal hardware accelerator configurations.
- Chemistry and Molecular Modeling: In the vast search space of molecular conformer generation, Chan et al. et al. utilize structural priors derived from existing databases to bias the initial surrogate model, finding low-energy shapes with 100x fewer evaluations than uniform searches.
By framing human intuition not as a rigid constraint, but as a flexible probabilistic bias, this research direction bridges the gap between human expertise and automated discovery—making black-box optimization drastically faster while preserving its mathematical guarantees.
Key Foundational Work
Automatic Gait Optimization with Gaussian Process Regression (2007) — contributed 1 concept elements to this direction.
Annealing Techniques For Unsupervised Statistical Language Learning (2004) — contributed 1 concept elements to this direction.
Multi-start Method with Prior Learning for Image Registration (2007) — contributed 1 concept elements to this direction.