Dynamic Multi-Objective Optimization

• DMOP is a framework where objective functions, constraints, and feasible sets change over time to capture shifting Pareto-optimal solutions.
• It uses rigorous mathematical models and dynamic elements, including time-dependent parameters and state-based dynamics, to handle fluctuating conditions.
• Applications span wireless networks, transportation, and scheduling, leveraging prediction, transfer learning, and diversity management for optimal performance.

A Dynamic Multi-Objective Optimization Problem (DMOP) is an optimization framework in which the objective functions, constraints, and often the feasible set are explicitly time-dependent, requiring continuous or event-driven re-optimization to track a moving set of Pareto-optimal solutions and corresponding Pareto fronts. DMOPs model systems subject to environmental or internal changes, including fluctuating demands, time-varying operational parameters, and shifting multi-criteria trade-offs.

1. Mathematical Formulation

A general DMOP at time $t$ is defined as

$$\min_{x \in \Omega} \quad F(x, t) = \left(f_1(x, t), f_2(x, t), \dots, f_{m(t)}(x, t)\right)^\top$$

subject to

$$g_i(x, t) \leq 0 \quad (i=1, \dots, n_g), \qquad h_j(x, t) = 0 \quad (j=1, \dots, n_h)$$

where:

• $x \in \Omega \subseteq \mathbb{R}^n$ denotes the decision variables.
• $m(t)$ may vary with $t$; that is, the number of objectives can change over time (Chen et al., 2016, Ruan et al., 2023).
• $f_i(x, t)$, $g_i(x, t)$, $h_j(x, t)$ are all explicit functions of time.

The dynamic Pareto set at time $t$ is

$$\min_{x \in \Omega} \quad F(x, t) = \left(f_1(x, t), f_2(x, t), \dots, f_{m(t)}(x, t)\right)^\top$$0

where $$\min_{x \in \Omega} \quad F(x, t) = \left(f_1(x, t), f_2(x, t), \dots, f_{m(t)}(x, t)\right)^\top$$1 denotes Pareto dominance at time $$\min_{x \in \Omega} \quad F(x, t) = \left(f_1(x, t), f_2(x, t), \dots, f_{m(t)}(x, t)\right)^\top$$2. The associated dynamic Pareto front is $$\min_{x \in \Omega} \quad F(x, t) = \left(f_1(x, t), f_2(x, t), \dots, f_{m(t)}(x, t)\right)^\top$$3 (Hou et al., 2023, Chen et al., 2016).

2. Taxonomy of Dynamic Elements and Model Classes

DMOPs are classified by the type and structure of time-dependency, following a unified taxonomy (Tantar et al., 2011, Herring et al., 2022, Shao et al., 4 Jan 2026):

1. Time-dependent parameters (input space): The mapping $$\min_{x \in \Omega} \quad F(x, t) = \left(f_1(x, t), f_2(x, t), \dots, f_{m(t)}(x, t)\right)^\top$$4 changes via transformations $$\min_{x \in \Omega} \quad F(x, t) = \left(f_1(x, t), f_2(x, t), \dots, f_{m(t)}(x, t)\right)^\top$$5.
2. Time-dependent objectives (output space): $$\min_{x \in \Omega} \quad F(x, t) = \left(f_1(x, t), f_2(x, t), \dots, f_{m(t)}(x, t)\right)^\top$$6 directly depends on $$\min_{x \in \Omega} \quad F(x, t) = \left(f_1(x, t), f_2(x, t), \dots, f_{m(t)}(x, t)\right)^\top$$7; the structure or shape of the objectives fluctuates.
3. State-dependent dynamics: Past states $$\min_{x \in \Omega} \quad F(x, t) = \left(f_1(x, t), f_2(x, t), \dots, f_{m(t)}(x, t)\right)^\top$$8 influence $$\min_{x \in \Omega} \quad F(x, t) = \left(f_1(x, t), f_2(x, t), \dots, f_{m(t)}(x, t)\right)^\top$$9 via memory or lag mechanisms.
4. Environmental/process-driven changes: The definitions, domains, constraints, or number of objectives themselves change due to exogenous events, stochastic disturbances, or regime switching (Chen et al., 2016, Ruan et al., 2023).

Dynamics can be periodic or aperiodic, with deterministic or random change frequencies and severities (Herring et al., 2022, Shao et al., 4 Jan 2026). In practice, the time index is often discretized as $$g_i(x, t) \leq 0 \quad (i=1, \dots, n_g), \qquad h_j(x, t) = 0 \quad (j=1, \dots, n_h)$$0 with $$g_i(x, t) \leq 0 \quad (i=1, \dots, n_g), \qquad h_j(x, t) = 0 \quad (j=1, \dots, n_h)$$1 as the generation counter, $$g_i(x, t) \leq 0 \quad (i=1, \dots, n_g), \qquad h_j(x, t) = 0 \quad (j=1, \dots, n_h)$$2 as the interval between changes, and $$g_i(x, t) \leq 0 \quad (i=1, \dots, n_g), \qquad h_j(x, t) = 0 \quad (j=1, \dots, n_h)$$3 encoding severity (Lei et al., 2024, Shao et al., 4 Jan 2026).

3. Benchmarks, Pareto Sets, and Dynamic Features

Benchmark DMOPs incorporate a variety of landscape dynamics:

• Moving or morphing Pareto sets: The PS can shift, rotate, deform, or expand/contract on manifolds of varying dimensionality, not restricted to hyperplanes (Shao et al., 4 Jan 2026, Chen et al., 2016, Ruan et al., 2023).
• Variable/expanding objective count: Changes in $$g_i(x, t) \leq 0 \quad (i=1, \dots, n_g), \qquad h_j(x, t) = 0 \quad (j=1, \dots, n_h)$$4 induce expansion or contraction of PF/PS. For instance, an increase in $$g_i(x, t) \leq 0 \quad (i=1, \dots, n_g), \qquad h_j(x, t) = 0 \quad (j=1, \dots, n_h)$$5 leads to $$g_i(x, t) \leq 0 \quad (i=1, \dots, n_g), \qquad h_j(x, t) = 0 \quad (j=1, \dots, n_h)$$6, and a decrease yields $$g_i(x, t) \leq 0 \quad (i=1, \dots, n_g), \qquad h_j(x, t) = 0 \quad (j=1, \dots, n_h)$$7 (Chen et al., 2016, Ruan et al., 2023).
• Time-linkage and error accumulation: Some problems encode time-coupling, such that suboptimal decision-making at time $$g_i(x, t) \leq 0 \quad (i=1, \dots, n_g), \qquad h_j(x, t) = 0 \quad (j=1, \dots, n_h)$$8 degrades the attainable PS and PF at $$g_i(x, t) \leq 0 \quad (i=1, \dots, n_g), \qquad h_j(x, t) = 0 \quad (j=1, \dots, n_h)$$9 due to "landscape pollution" (Shao et al., 4 Jan 2026).
• Controlled variable dominance and rotation matrices: Variable weights and time-dependent non-separability control the heterogeneity and hardness of the problem space (Shao et al., 4 Jan 2026).

This diversity is captured in advanced test suites (e.g., GTS (Shao et al., 4 Jan 2026), DTLZ/WFG variants (Chen et al., 2016, Ruan et al., 2023)) and emphasized in reproducibility and benchmarking frameworks (Herring et al., 2022), which recommend invariance and systematic reporting across severity/frequency parameters.

4. Solution Strategies and Algorithmic Architectures

Algorithmic approaches to DMOPs are specialized extensions of evolutionary multi-objective algorithms (EMOAs) and typically exhibit the following features:

4.1 Prediction and Knowledge Transfer

• Feed-forward models: Autoregressive, centroid-shift, or kernel-based prediction models anticipate movement of PF/PS, using recent historical data to seed the next population (Hou et al., 2023, Lei et al., 2024).
• Transfer learning and regression: Regressors built on source-target pairs or boosting ensembles are used to map previous populations onto the new front following environmental change (Wang et al., 2019).
• Support Vector Machines (SVM/ISVM): Classification of good/bad regions based on the last POS provides filtered initial populations in new environments (Jiang et al., 2019, Hu et al., 2019).

4.2 Archive and Diversity Management

• Multi-archive methods: Two-archive models separate convergence-driven and diversity-driven populations, reconstructed adaptively upon dynamic events (Chen et al., 2016).
• Knowledge-transfer operators: For changing $x \in \Omega \subseteq \mathbb{R}^n$0, explicit expansion/contraction operators transfer the PS between different-dimensional PFs, producing a spread of non-dominated solutions that immediately recover diversity and coverage (Ruan et al., 2023).

4.3 Change Detection and Response

• Detectors: Random or archive-based re-evaluations identify changes, triggering re-initialization or prediction pipelines (Hou et al., 2023, Jiang et al., 2019, Aboud et al., 2019).
• Controlled response: Strategies include partial reinitialization, memory-based elitism, adaptive mutation, or complete random restart, each with characteristic effects on convergence and tracking (Herring et al., 2022, Li et al., 2022).

4.4 Hierarchical and Hybrid Frameworks

• Layered schemes: Local feature aggregation and global adjustment provide synergies for high-dimensional dynamic network optimizations (Ren et al., 21 Jan 2026).
• Adaptive prediction weighting: Dual-space (decision/objective) predictors are adaptively combined based on current population informativeness (Lei et al., 2024).

5. Objective Functions, Constraints, and Analytical Models

DMOPs exhibit tightly coupled objectives with analytical evaluation models:

• Coverage, Overlap, Spectral Efficiency, Power Consumption: In wireless systems, objectives are computed via analytical models that combine radio propagation, stochastic geometry, and traffic field integration (Chandhar et al., 2016).
• Normalization and dynamic weighting: Objectives are often normalized or adaptively weighted according to system urgency or real-time criteria (Ren et al., 21 Jan 2026).
• Feasibility constraints: Constraints may enforce blocking probability bounds, minimum coverage, or resource/parameter intervals. They are strictly imposed in fitness evaluation (Chandhar et al., 2016).
• Composite metrics: Trade-off surfaces (Pareto fronts) are constructed, with operators selecting optimal points according to operational policy (e.g., energy saving vs. capacity premium) (Chandhar et al., 2016).

6. Performance Metrics and Empirical Evaluation

Standard metrics for tracking convergence and diversity in DMOPs include:

Metric	Mathematical Expression	Interpretation
Inverted Generational Distance (IGD)	$x \in \Omega \subseteq \mathbb{R}^n$1	Convergence to the true PF at time $x \in \Omega \subseteq \mathbb{R}^n$2
Hypervolume (HV)	$x \in \Omega \subseteq \mathbb{R}^n$3	Coverage/diversity in objective space
Maximum Spread (MS)	$x \in \Omega \subseteq \mathbb{R}^n$4, $x \in \Omega \subseteq \mathbb{R}^n$5	Coverage of the PF extremities

Averaged over time, dynamic MIGD ($x \in \Omega \subseteq \mathbb{R}^n$6), MHV ($x \in \Omega \subseteq \mathbb{R}^n$7), and DMIGD/DMHV benchmark entire DMOP sequences and facilitate statistical comparisons across methods (Shao et al., 4 Jan 2026, Chen et al., 2016, Hou et al., 2023, Ruan et al., 2023).

7. Application Domains and Practical Impact

DMOPs have been instantiated in several high-impact domains:

• Green wireless network optimization: Dynamic selection of active sectors, transmit powers, and antenna parameters maximizes coverage and spectral efficiency while minimizing overlap and area power consumption, subject to stochastic load variations and interference models (Chandhar et al., 2016).
• Topology optimization in vehicular networks: Dynamic adaptation of communication links and bandwidth allocations balances latency, throughput, and path redundancy under node mobility and connectivity constraints (Ren et al., 21 Jan 2026).
• Resource allocation, transportation, scheduling: Time-evolving multi-criteria trade-offs require continuous adjustment of operational parameters to maintain optimality under uncertainty (Tantar et al., 2011, Shao et al., 4 Jan 2026).

Consistent empirical results show that dynamically sophisticated EMOAs (prediction, transfer learning, knowledge-driven population initialization) provide marked improvements over static or naive baselines, especially under high-severity changes, fast-changing environments, or variable objective cardinality (Chen et al., 2016, Ruan et al., 2023, Hou et al., 2023, Lei et al., 2024).

---

References:

• (Chandhar et al., 2016) Multi-Objective Framework for Dynamic Optimization of OFDMA Cellular Systems
• (Chen et al., 2016) Dynamic Multi-Objectives Optimization with a Changing Number of Objectives
• (Hou et al., 2023) Combining Kernelized Autoencoding and Centroid Prediction for Dynamic Multi-objective Optimization
• (Ruan et al., 2023) Knowledge Transfer for Dynamic Multi-objective Optimization with a Changing Number of Objectives
• (Lei et al., 2024) An accelerate Prediction Strategy for Dynamic Multi-Objective Optimization
• (Herring et al., 2022) Reproducibility and Baseline Reporting for Dynamic Multi-objective Benchmark Problems
• (Shao et al., 4 Jan 2026) Benchmarking Continuous Dynamic Multi-Objective Optimization: Survey and Generalized Test Suite
• (Jiang et al., 2019) Solving dynamic multi-objective optimization problems via support vector machine
• (Hu et al., 2019) Solving Dynamic Multi-objective Optimization Problems Using Incremental Support Vector Machine
• (Wang et al., 2019) Evolutionary Dynamic Multi-objective Optimization Via Regression Transfer Learning
• (Tantar et al., 2011) Design and classification of dynamic multi-objective optimization problems
• (Ren et al., 21 Jan 2026) Hierarchical Optimization Based Multi-objective Dynamic Regulation Scheme for VANET Topology
• (Li et al., 2022) A Framework Based on Generational and Environmental Response Strategies for Dynamic Multi-objective Optimization
• (Aboud et al., 2019) Dynamic Multi Objective Particle Swarm Optimization based on a New Environment Change Detection Strategy