Dynamic Multi-Objective Optimization under Uncertainty

• Dynamic multi-objective optimization under uncertainty is a framework that computes Pareto-optimal solutions by accounting for variable parameters in both objectives and constraints.
• It employs robust, set-based, and worst-case formulations, integrating Bayesian and reinforcement learning methods to address diverse forms of uncertainty.
• Advanced algorithms like quasi-Newton, Bayesian optimization, and multi-stage scenario-based approaches enable practical applications in engineering, finance, and supply chain management.

Dynamic multi-objective optimization under uncertainty addresses the problem of computing or approximating sets of Pareto-optimal or robustly optimal solutions for vector-valued objectives when parameters in objectives and/or constraints are subject to uncertainty. The problem is critical in domains where multiple conflicting criteria must be balanced under incomplete or unreliable information—examples include engineering design, finance, supply chain management, and policy optimization. The field builds on mathematical programming, control theory, and stochastic optimization, deploying techniques ranging from robust and set-based minmax formulations to Bayesian surrogate modeling, scenario-based planning, and reinforcement learning. Approaches must account for various forms and sources of uncertainty: parametric, structural, and exogenous, both in dynamic (time-dependent) and static environments.

1. Mathematical Formulations of Robust Multi-Objective Optimization

Several formulations of robustness and efficiency have emerged for multi-objective optimization under uncertainty.

Highly Robust Solutions:

A solution $x^* \in \Omega$ is highly robust (weakly) efficient if it is efficient for all possible uncertainty scenarios $u \in \mathcal{U}$:

$$x^* \text{ is highly robust efficient} \Longleftrightarrow \forall u \in \mathcal{U},\; x^* \text{ is efficient for } f(x,u).$$

This strict requirement produces solutions that are Pareto-efficient across all realizations of uncertain parameters (Rahimi et al., 11 Jan 2025).

Set-Based and Minmax Approaches:

In the set-based minmax framework, robustness is defined via set-valued objectives:

$$F_{\mathcal{U}}(x) = \{ f(x,u): u \in \mathcal{U} \}.$$

Efficiency is established with respect to the strict upper set ordering, stating that no feasible alternative $x'$ yields an image $F_{\mathcal{U}}(x')$ that dominates $F_{\mathcal{U}}(x)$ throughout the uncertainty set (Eichfelder et al., 2022).

Worst-Case and Cardinality-Constrained Robustness:

Minmax formulations replace each objective with the supremum (or infimum for cost minimization) over specified uncertainty sets. For example, with finite discrete scenarios $U = \{\xi_1, ..., \xi_N\}$, each robust objective becomes

$F_j(x) = \max_{i} f_j(x, \xi_i).$

This yields a deterministic multi-objective optimization with nonsmooth worst-case objectives (Kumar et al., 2023, Gupta et al., 20 May 2025).

Scenario- and Conditional Pareto Analysis:

The conditional Pareto front $P(u)$ is defined for each scenario $u$, and robustness is quantified via the coverage probability:

$\mathbb{P}_U[x \in P(U)],$

interpreted as the probability that $x$ is Pareto-optimal under realization $u$ (Trappler et al., 7 Apr 2025).

2. Solution Characterizations and Optimality Conditions

Optimality conditions for robust solutions extend classical KKT theory via variational analysis and subdifferential calculus.

Necessary and Sufficient Conditions:

For locally Lipschitz $f_i$, a local highly robust (weakly) efficient $x^*$ satisfies:

$$\forall\, d \in T_w(x^*;\Omega),\; \forall\, u \in \mathcal{U},\; \exists\, i:\;\langle x^* - u_i, d \rangle \geq 0,\;\text{for some } x^* \in \partial f_i(x^*).$$

Sufficiency can be established under additional generalized convexity assumptions (Rahimi et al., 11 Jan 2025).

Set and Vectorization Approaches:

Set-based robust optimization allows reformulation into standard multi-objective problems via vectorization. In (Eichfelder et al., 2022), images are regularized to preserve compactness and weakly efficient robust solutions converge as $p \to \infty$ in the vectorized multiobjective problem:

$$\operatorname{min}_{(x, y^1, ..., y^p)} (y^1, ..., y^p)\;\;\text{subject to}\;\; y^i \in F(x),\; x \in \Omega.$$

Quasi-Newton and Descent Algorithms:

For robust counterparts under finite scenario uncertainty, quasi-Newton methods construct locally quadratic models for each scenario and objective, updating Hessian approximations (often via a BFGS-type formula) and relying on partition set decompositions to organize the possible maximal active indices in robust set-valued images (Kumar et al., 2023, Gupta et al., 20 May 2025). Line search and convergence analysis rely on scalarization via Gerstewitz functionals and Armijo conditions.

3. Classes of Uncertainty Sets and Their Impact

The form of the uncertainty set $\mathcal{U}$ critically affects both practical tractability and solution conservatism.

• Ball and Ellipsoidal Sets:

If $0 \in \operatorname{int} \mathcal{U}_i$ for all $i$, robust weakly efficient solutions coincide with strict efficiency in the nominal problem, due to the ability of $u$ to perturb in any direction (Rahimi et al., 11 Jan 2025).

• Polyhedral Sets:

When $\mathcal{U}_i$ is polyhedral, robust optimality can be checked based solely on finitely many extreme points and rays of $\mathcal{U}_i$, simplifying verification and computation (Rahimi et al., 11 Jan 2025).

• Box, Norm, and Ellipsoidal Uncertainty in Constraints:

For uncertain linear and conic constraints, tractability is achieved since robust feasibility and efficiency checks reduce to finite constraint verification—e.g., for norm or ellipsoidal uncertainty, robust feasibility/scalability can be verified via systems of conic inequalities (Goberna et al., 2014).

• Cardinality-Constrained Uncertainty:

Robust combinatorial optimization under $K$-cardinality uncertainty is tractable via subproblem decompositions, sum and bottleneck function reformulations, and polynomial-time label-setting algorithms (Raith et al., 2017).

4. Algorithmic and Computational Methods

Set Optimization and Partitioned Descent:

Algorithms for robust (weakly) efficient solutions in set optimization proceed by decomposing the robust image $F_U(x)$ via partition sets, optimizing via scalarizations, and using BFGS/quasi-Newton updates in vector optimization subroutines (Gupta et al., 20 May 2025). Global convergence (to stationary points under upper set less ordering) and local superlinear rates are provable under uniform continuity and regularity.

Bayesian Optimization under Uncertainty:

Gaussian Process-based frameworks robustify objectives by modeling the Bayes risk (the expected value under input noise or environmental parameters). Uncertainty is integrated via Monte Carlo or analytical kernel expectations; acquisition functions for batch candidate selection are extended to incorporate the volume of uncertainty hyper-rectangles or quantile improvement, and active learning steps mitigate regions of persistent uncertainty (Qing et al., 2022, Belakaria et al., 2022, Belakaria et al., 2020, Semochkina et al., 22 Jan 2024, Inatsu et al., 2023, Trappler et al., 7 Apr 2025).

Active Learning and Coverage Estimation:

Conditional Pareto front analysis is supported by surrogate models and active learning to estimate coverage probabilities accurately, focusing evaluations via reformulated Expected Hypervolume Improvement (EHVI) integrated over uncertainty (Trappler et al., 7 Apr 2025).

Scenario-Based and Multi-stage Approaches:

In problems of deep uncertainty where no reliable probability measures are available, robust adaptive strategies employ multi-stage, multi-scenario optimization: meta-decisions are planned across scenario trees, and Pareto analysis is performed over all scenario paths, often employing reference point goal programming for practical decision support (Shavazipour et al., 2023).

5. Risk-Based and Quantitative Measures of Uncertainty

Risk-Measure Robust Optimization:

Several approaches robustify the multi-objective function with respect to risk measures (e.g., Bayes risk, Value-at-Risk, CVaR). This allows for direct control of the tail behavior of each objective's distribution under uncertain inputs. Bounding box-based MOBO methods construct high-probability intervals for risk measures and select query points maximizing the discrepancy between upper and lower risk bounds, with finite-sample theoretical guarantees on convergence accuracy (Inatsu et al., 2023, Kotecha et al., 8 Sep 2025).

Quantifying the Cost of Uncertainty:

The mean multi-objective cost of uncertainty (MOCU) evaluates the expected increase in operational cost due to model uncertainty, integrating over both uncertainty in model parameters and weighting of multiple objectives. This provides a principled metric for optimal experiment or data acquisition design, tractably quantifying degradation due to uncertainty on goal-specific objectives (Yoon et al., 2020).

6. Practical Applications and Domains

Multifaceted robust multi-objective optimization methods have found concrete applications in:

• Engineering and Circuit Design: Optimal design of switched-capacitor voltage regulators and cabin insulation under manufacturing and environmental uncertainty employ surrogate-based BO, Pareto front approximation under constraint satisfaction, and active learning to reduce expensive evaluations (Belakaria et al., 2020, Trappler et al., 7 Apr 2025).
• Supply Chain Management: Multi-objective reinforcement learning (with evolutionary optimization and risk measures like CVaR) provides adaptable, diverse policies for inventory and logistics, allowing dynamic switching between strategies as system objectives or constraints shift (Kotecha et al., 8 Sep 2025).
• Markov Decision Processes: Policy synthesis in MDPs with uncertain transition rates associates each policy with worst-case, best-case, and expected vector performance, using Pareto optimality to inform flexible and robust decision-making (Scheftelowitsch et al., 2017, Chen et al., 2021).
• Portfolio Optimization: Scenario-based and multi-stage robust optimization methods deliver adaptive and dynamically robust strategies that maintain satisfactory returns and consumption across plausible future market paths, without requiring reliable probabilities (Shavazipour et al., 2023, Rahimi et al., 11 Jan 2025).
• Public Health and Environmental Policy: Multi-objective BO methods identify robust intervention strategies accounting for uncertainty in environmental dynamics, as in disease outbreak management with environmental dispersion models (Semochkina et al., 22 Jan 2024).

7. Comparative Analysis and Ongoing Research Directions

Trade-offs and Method Selection:

The degree of conservatism, computational tractability, and generality differs widely between highly robust, set-based, worst-case, scenario-based, and probabilistic approaches. Highly robust efficiency is the strongest (most conservative) notion but may produce overly restrictive solutions if uncertainty sets are large. Set and vectorization methods offer a balance between computational tractability and solution generality (Eichfelder et al., 2022, Rahimi et al., 11 Jan 2025).

Algorithmic Efficiency:

Quasi-Newton and Newton-type schemes with BFGS updates are competitive on nonsmooth, nonconvex deterministic robust counterparts but require strong differentiability. Surrogate-based Bayesian approaches, while more flexible with black-box objectives or uncertain constraints, can exploit parallel computation and active learning to address high-dimensional or expensive scenarios (Qing et al., 2022).

Emerging Techniques:

Recent work integrates human-in-the-loop preference elicitation, interpretable and permissive control synthesis, reinforcement learning with evolutionary population-based search, and scenario-based adaptive planning for deep uncertainty (Chen et al., 2021, Kotecha et al., 8 Sep 2025, Shavazipour et al., 2023). Risk-aware stochastic optimization (e.g., CVaR or distributionally robust MOBO) continues to gain traction for its ability to control for rare catastrophic outcomes.

A plausible implication is that future research will increasingly focus on scalable mixed-integer, surrogate-based, and learning-augmented methods that combine scenario, set, and risk-based robustness, with special attention to integration in real-time automation and human-in-the-loop decision support systems.

---

In summary, dynamic multi-objective optimization under uncertainty is a multidisciplinary research arena characterized by rigorous mathematical frameworks (set-based, minmax, scenario, and Bayesian), powerful solution algorithms (quasi-Newton, Bayesian optimization, MOEAs, RL), and a broad spectrum of real-world applications. Theoretical advances in optimality conditions, tractable reformulations, and uncertainty quantification have significantly advanced both exact and approximate solution approaches, with current trends accelerating adoption of hybrid, risk-averse, and adaptive methods in complex, data-driven environments.