Abstract
Bayesian optimization (BO) is a popular approach for sample-efficient optimization of black-box objective functions. While BO has been successfully applied to a wide range of scientific applications, traditional approaches to single-objective BO only seek to find a single best solution. This can be a significant limitation in situations where solutions may later turn out to be intractable. For example, a designed molecule may turn out to violate constraints that can only be reasonably evaluated after the optimization process has concluded. To address this issue, we propose Rank-Ordered Bayesian Optimization with Trust-regions (ROBOT) which aims to find a portfolio of high-performing solutions that are diverse according to a user-specified diversity metric. We evaluate ROBOT on several real-world applications and show that it can discover large sets of high-performing diverse solutions while requiring few additional function evaluations compared to finding a single best solution.
Introduction
In high-stakes domains such as drug discovery, materials science, and engineering, the "optimal" solution identified by an optimization algorithm often fails when transitioned to real-world testing. A candidate molecule might show exceptional antibiotic activity in silico but prove toxic in clinical trials. If a Bayesian Optimization (BO) process yields only a single global optimum, the failure of that candidate renders the entire computational effort moot.
Traditional BO is essentially an "all-or-nothing" strategy. While some variants attempt to find multiple solutions, they often rely on simple Euclidean distance in the input space to define diversity. This paper introduces ROBOT (Rank-Ordered Bayesian Optimization with Trust-regions), a framework designed to discover a portfolio of high-performing, diverse solutions based on arbitrary, user-defined diversity metrics.
Methodology: The ROBOT Framework
ROBOT addresses the diversity challenge by treating it as a constraint rather than a secondary objective. The architecture is built upon three primary components:
- Multiple Trust Regions ($M$): ROBOT maintains $M$ local trust regions (hyper-rectangles) within the search space. Each region is tasked with identifying one high-quality solution within the portfolio.
- Hierarchical Rank-Ordering: To ensure diversity without destabilizing the search, trust regions are ranked from 1 to $M$. The top-ranked region ($\mathcal{T}_1$) performs an unconstrained search for the global optimum. Each subsequent region $\mathcal{T}_i$ is constrained to find a solution that is diverse (meeting a threshold $\tau$) from all solutions found by higher-ranked regions ($\mathcal{T}_1, \dots, \mathcal{T}_{i-1}$). This asymmetric structure prevents lower-ranked regions from interfering with the search for the global optimum.
- Global Surrogate Model: Unlike previous multi-trust-region methods that utilize isolated local models, ROBOT employs a single Parametric Gaussian Process Regressor (PPGPR). This global model allows all trust regions to share evaluation data, significantly improving sample efficiency.
For high-dimensional or structured data, such as molecular graphs, the authors propose LOL-ROBOT, which operates within the latent space of a Variational Autoencoder (VAE).
Experimental Evaluation
The researchers benchmarked ROBOT against several state-of-the-art baselines across diverse tasks including Rover Trajectory Optimization, Lunar Lander Policy, and Molecular Design (using the Guacamol benchmark).
Key Results
ROBOT demonstrated superior performance in discovering large sets of high-quality solutions. In molecular tasks, it consistently outperformed Standard BO and Graph-GA, identifying significantly more high-scoring, diverse molecules. Notably, the "cost" of finding a portfolio is not linear; ROBOT often discovers dozens of solutions with only a 3x to 6x increase in function evaluations compared to finding one.
Furthermore, the Lunar Lander experiments highlighted the practical value of diversity. An ensemble of 20 diverse policies generated by ROBOT proved far more robust to environmental variations than a single "optimal" policy, which was prone to catastrophic failure in unseen conditions.
Discussion and Conclusion
ROBOT shifts the paradigm of BO from finding a single point to discovering a robust portfolio. By treating diversity as a threshold constraint and utilizing a hierarchical ranking system, the method avoids the complexities of multi-objective optimization while maintaining high sample efficiency through a shared global surrogate.
Limitations: The current framework assumes that the diversity metric is computationally inexpensive to evaluate. Additionally, the user must pre-specify the diversity threshold $\tau$, which requires domain knowledge. Future work may explore defining trust regions directly within the diversity metric space to further enhance search efficiency in non-Euclidean domains.
Summary updated