Abstract
We present Parallel Feasible Pareto Frontier Entropy Search ($\text{PF}^2$ES) — a novel information-theoretic acquisition function for multi-objective Bayesian optimization supporting unknown constraints and batch query. Due to the complexity of characterizing the mutual information between candidate evaluations and (feasible) Pareto frontiers, existing approaches must either employ crude approximations that significantly hamper their performance or rely on expensive inference schemes that substantially increase the optimization's computational overhead. By instead using a variational lower bound, $\text{PF}^2$ES provides a low-cost and accurate estimate of the mutual information. We benchmark $\text{PF}^2$ES against other information-theoretic acquisition functions, demonstrating its competitive performance for optimization across synthetic and real-world design problems.
Introduction
Multi-objective optimization (MOO) is a cornerstone of engineering and machine learning, where the goal is to find optimal trade-offs—the Pareto frontier—between conflicting objectives, such as a vehicle's fuel efficiency versus its safety. When objective functions are computationally expensive to evaluate, Bayesian Optimization (BO) provides a sample-efficient framework by utilizing a surrogate model, typically a Gaussian Process (GP), and an acquisition function to guide the search.
However, existing information-theoretic acquisition functions often struggle with the mathematical complexity of calculating mutual information between samples and the Pareto frontier. Furthermore, real-world applications frequently involve unknown constraints and require batch (parallel) evaluations, areas where current methods are often inefficient or rely on crude approximations.
The {PF}$^2$ES Framework
The authors introduce Parallel Feasible Pareto Frontier Entropy Search ({PF}$^2$ES), a novel acquisition function designed to overcome these challenges using a variational lower bound for efficient and accurate mutual information estimation.
Key Technical Innovations
- Output Space Partitioning: {PF}$^2$ES partitions the objective space into dominated and non-dominated regions through hypervolume-based decomposition. This allows for a structured approach to estimating the entropy of the Pareto frontier.
- Handling Unknown Constraints: The method explicitly incorporates feasibility by considering the joint distribution of objectives and constraints, ensuring that the search focuses on the feasible Pareto frontier.
- $\epsilon$-Penalization for Robustness: To prevent the algorithm from over-exploiting known regions (clustering), a small penalization vector ($\epsilon$) is introduced to shift the frontier, encouraging broader exploration.
- Parallelization ($q$-{PF}$^2$ES): The framework is extended to support batch queries of size $q$ using Monte Carlo approximations and the reparameterization trick, making it suitable for modern parallel computing environments.
Experimental Evaluation
The performance of {PF}$^2$ES was benchmarked against state-of-the-art methods like Expected Hypervolume Improvement (EHVI) and Multi-Objective Max-value Entropy Search (MESMO) across a variety of tasks, including standard synthetic functions (VLMOP2, ZDT, DTLZ4) and real-world engineering problems such as Four Bar Truss design and Disc Brake design.
Key Results
- State-of-the-Art Performance: {PF}$^2$ES consistently matched or outperformed established baselines in terms of convergence speed and final hypervolume across almost all tests.
- No Reference Point Needed: Unlike the "gold standard" EHVI, {PF}$^2$ES does not require a pre-defined reference point, which is a major practical advantage when the range of possible outcomes is unknown.
- Scalability and Robustness: The method demonstrated strong performance in batch settings and showed low sensitivity to internal hyperparameters, making it a robust "plug-and-play" solution for researchers.
Discussion and Limitations
While {PF}$^2$ES represents a significant step forward, its performance remains dependent on the quality of the underlying GP surrogate. Additionally, while the partitioning scheme is efficient, its computational cost still scales with the number of objectives ($M$). Future research may explore more fully Bayesian approaches for final frontier recommendations and scaling to even larger batch sizes.
Summary updated