Multi-Objective Bayesian Optimization (MOBO)

MOBO uses a probabilistic model, typically a Gaussian Process (GP), to approximate the expensive, unknown objective functions. This model provides a predictive distribution (mean and variance) for each objective at any untested point in the search space. The key advantage is sample efficiency; the model quantifies uncertainty, allowing the algorithm to make informed decisions with very few direct evaluations of the costly true functions.

The core decision engine of MOBO is the acquisition function, which uses the surrogate model's predictions to score candidate points for evaluation. Unlike single-objective BO, MOBO acquisition functions must balance exploration, exploitation, and Pareto front discovery. Common strategies include:

• Expected Hypervolume Improvement (EHVI): Measures the expected increase in the dominated volume of the objective space.
• ParEGO: Applies scalarization (e.g., the weighted Chebyshev method) to the GP's predictions.
• Probability of Improvement-based metrics for multiple objectives.

The primary output of a MOBO run is an approximation of the Pareto front—the set of optimal trade-off solutions. The algorithm iteratively selects evaluation points expected to either:

• Expand the known front by discovering new, non-dominated regions.
• Fill in gaps to improve the density and coverage of the front.
• Refine solutions near areas of particular interest (if preferences are given). The final set of evaluated points provides the decision-maker with a clear visualization of the available trade-off surface.

MOBO is explicitly designed for problems where evaluating the objective functions is prohibitively costly in terms of time, money, or computational resources. Examples include:

• Hyperparameter tuning for deep learning models (balancing accuracy, latency, model size).
• Engineering design (e.g., aerodynamic shape optimization balancing drag, lift, and structural stress).
• Scientific simulation (e.g., drug discovery balancing potency, selectivity, and synthetic accessibility). By strategically choosing the most informative points to evaluate, MOBO finds high-quality Pareto fronts in tens to hundreds of evaluations, where grid or random search would require thousands.

MOBO frameworks can incorporate preference articulation to focus the search on relevant regions of the Pareto front. This moves beyond simply finding the entire front to finding solutions that match specific stakeholder goals. Methods include:

• Reference point methods: Guiding the search towards a desired region defined by aspiration levels.
• Preference-based acquisition functions: Modifying the acquisition function to favor improvements aligned with stated preferences.
• Interactive MOBO: Allowing the decision-maker to provide feedback during the optimization loop, refining the target region iteratively.

MOBO differs fundamentally from population-based methods like Multi-Objective Evolutionary Algorithms (MOEAs) such as NSGA-II. While MOEAs excel at exploring complex, discontinuous spaces and can handle many objectives, they typically require thousands to millions of function evaluations. MOBO's strength is its extreme sample efficiency for expensive problems. It is often used in a hybrid fashion, where a MOEA performs an initial broad exploration, and MOBO performs a final, intensive refinement of the most promising regions.

The Pareto front (or Pareto frontier) is the set of all Pareto optimal solutions plotted in the objective space. It visually represents the best possible trade-offs between competing objectives; improving one objective necessarily worsens another along this frontier. In MOBO, the goal is to efficiently sample and approximate this high-dimensional surface.

• Key Property: Defines the optimal trade-off boundary.
• Visualization: A curve (2 objectives) or surface (3+ objectives).
• MOBO Goal: The surrogate model aims to predict and guide exploration towards this frontier.

A solution is Pareto optimal if no objective can be improved without degrading at least one other objective. It is a state of resource allocation efficiency. In MOBO, we seek a diverse set of Pareto optimal points to map the trade-offs.

• Formal Definition: A solution x* is Pareto optimal if there is no other solution x such that f_i(x) ≤ f_i(x*) for all objectives i and f_j(x) < f_j(x*) for at least one j.
• Implication: There is no single "best" solution, only a set of incomparable optimal trade-offs.

Scalarization transforms a multi-objective problem into a single-objective problem by aggregating the vector of objectives into a scalar. This is a common technique within MOBO frameworks like MOEA/D.

• Weighted Sum Method: Combines objectives: F_scalar = w1*f1 + w2*f2 + ... + wk*fk. Different weight vectors yield different points on the Pareto front.
• Epsilon-Constraint Method: Optimizes one primary objective while constraining others: min f1(x) subject to f2(x) ≤ ε2, f3(x) ≤ ε3, ...
• Role in MOBO: Enables the use of standard Bayesian Optimization by creating a scalar acquisition function from multiple objectives.

The hypervolume indicator (or S-metric) is a Pareto-compliant performance metric that measures the volume of the objective space dominated by a set of solutions, relative to a predefined reference point. It simultaneously rewards convergence towards the Pareto front and diversity of coverage.

• Calculation: The union of hyper-rectangles defined by each solution and the reference point.
• Use in MOBO: Often used as the basis for an acquisition function (e.g., Expected Hypervolume Improvement, EHVI) to directly guide the search towards maximizing the quality of the approximated Pareto set.

MOEAs are population-based metaheuristics (e.g., genetic algorithms) designed for multi-objective optimization. They contrast with MOBO's sample-efficient, model-based approach.

• Key Examples: NSGA-II (uses non-dominated sorting and crowding distance), MOEA/D (decomposes problem into scalar subproblems).
• Comparison to MOBO: MOEAs typically require 10,000+ function evaluations and are effective for cheap-to-evaluate functions. MOBO uses a surrogate model to reduce expensive evaluations to 100s.
• Hybrid Use: MOBO can seed or be combined with MOEAs for refinement.

In Bayesian Optimization, the acquisition function uses the surrogate model's predictions (mean and uncertainty) to decide the next most promising point to evaluate. In MOBO, it must balance exploring multiple objectives.

• Single-Objective Analogs: Expected Improvement (EI), Upper Confidence Bound (UCB).
• Multi-Objective Variants:
  • Expected Hypervolume Improvement (EHVI): Measures expected gain in hypervolume.
  • ParEGO: Applies scalarization (Tchebycheff) to the surrogate's prediction.
  • Probability of Improvement (PoI): Measures likelihood a point dominates a reference set.
• Purpose: Automates the trade-off between exploration (high uncertainty) and exploitation (good predicted mean).