Multi-Objective Genetic Algorithms

• Multi-objective genetic algorithms are population-based metaheuristics that optimize multiple competing objectives simultaneously.
• The literature distinguishes between Pareto-based, rank-based, and scalarized formulations, with methods such as NSGA-II demonstrating practical applications in diverse domains.
• MOO-GA employs domain-specific representations, tailored genetic operators, and advanced diversity and feasibility controls to address complex, real-world optimization problems.

Multi-objective genetic algorithm (MOO-GA) denotes a family of population-based optimization methods that evolve candidate solutions under multiple competing criteria rather than a single scalar objective. In the literature represented here, the term covers both Pareto-based evolutionary algorithms that preserve non-dominated trade-offs and scalarized genetic algorithms that aggregate several criteria into one fitness value before selection; accordingly, MOO-GA refers less to one fixed algorithm than to a design space of representations, objective formulations, and survivor-selection rules (0809.0416, Brookhouse et al., 2023).

1. Formal structure and problem classes

Across the cited work, MOO-GA is usually posed as optimization of an objective vector over a feasible set. One generic form is

$$\min_{x\in\Omega} F(x)=\bigl(f_1(x),f_2(x),\ldots,f_m(x)\bigr),$$

with Pareto-optimality defined by the absence of another feasible solution whose objective vector dominates it (Pan et al., 2016). In practice, this abstract form appears in markedly different decision spaces: fixed-cardinality subset selection for reference compound lists (Ohto et al., 11 May 2025), joint replica-count and placement optimization in container orchestration (Guerrero et al., 2024), binary feature-subset search for fair classification (Brookhouse et al., 2023), and route-set plus split-ratio optimization in decentralized exchange routing (Marfinetz, 24 Oct 2025).

The decision variables may be binary, integer, real-coded, or hybrid. In fair feature selection, each chromosome is a binary indicator over original features (Brookhouse et al., 2023). In toxicity reference-list design, the decision variable is a fixed-size subset of compounds (Ohto et al., 11 May 2025). In accelerator tuning, the variables are cavity RF gradients under bound and energy constraints (Rajput et al., 2024). In cloud allocation, the chromosome simultaneously encodes the number of containers for each microservice and the physical machines that host them (Guerrero et al., 2024). This breadth of modeling choices is one of the defining characteristics of MOO-GA practice.

The objective sets are equally heterogeneous. Some papers optimize conventional performance trade-offs, such as minimizing heat load and trip rate in particle accelerators (Rajput et al., 2024) or minimizing threshold distance, cluster imbalance, system failure, and total network distance in container allocation (Guerrero et al., 2024). Others optimize domain-specific combinations, such as structural, physicochemical, and toxicity diversity in compound-list design (Ohto et al., 11 May 2025), or cost inefficiency, service failure, and employee dissatisfaction in healthcare scheduling (Patel et al., 28 Aug 2025). This suggests that the generality of MOO-GA lies primarily in its search framework rather than in any standard objective vocabulary.

2. Pareto-based, rank-based, and scalarized formulations

A central distinction in the literature is between true Pareto-based MOO-GA and scalarized multi-criteria GA. Pareto-based methods rank solutions by dominance relations and preserve non-dominated trade-offs. NSGA-II is the most common engine in the supplied papers, appearing in container allocation, toxicity reference-list optimization, healthcare scheduling, accelerator tuning, and several other applications (Guerrero et al., 2024, Ohto et al., 11 May 2025, Patel et al., 28 Aug 2025, Rajput et al., 2024). NSGA-III appears in recurrent spiking-neural-network optimization, where solutions are sorted by non-dominance and then normalized, associated, and niche-preserved (Fitzgerald et al., 2021).

A simpler Pareto-based alternative is dominance-count ranking. In multi-objective vehicle routing, the fitness of individual $i$ is derived from $\xi_i$, the number of population members that dominate it, with monotonicity conditions

$$\text{if } \xi_i < \xi_j \rightarrow f(i) > f(j), \qquad \text{if } \xi_i = \xi_j \rightarrow f(i) = f(j),$$

and linear normalization between $f_{\max}$ and $f_{\min}$ (0809.0416). All nondominated individuals have $\xi_i=0$ and therefore tie for maximal fitness.

Several papers explicitly separate this Pareto tradition from weighted scalarization. In CNN pruning, the optimization target is

$f = \lambda_1 e + \lambda_2 c + \lambda_3 (1-s),$

and the paper explicitly states that the method is not a true Pareto MOEA because it uses a weighted aggregated scalar fitness rather than non-dominated sorting (Yang et al., 2019). A comparable pattern appears in music harmony generation, where grammar, predicted expert score, and predicted listener score are combined as

$\text{Fitness}(x)=W_1X_1(x)+W_2X_2(x)+W_3X_3(x)$

inside a single scalar objective (Majidi et al., 2021). The prey-predator hardness study likewise aggregates lifespan, challenge, and usability into

$F = L + C + U$

and ranks chromosomes by that scalar quantity (Ansari et al., 2014).

A third formulation is lexicographic optimization. In fair feature selection, the lexicographic GA gives higher priority to predictive accuracy and only invokes fairness comparisons when accuracy differences fall below a threshold $i$0; by contrast, the Pareto GA treats accuracy and fairness as co-optimized without fixed priority (Brookhouse et al., 2023). This comparison shows that MOO-GA can encode objective relations not only through Pareto dominance or weights, but also through ordered preference structures.

3. Representations and genetic operators

Chromosome design in MOO-GA is typically domain-specific rather than generic. In fair and multi-view feature selection, the representation is binary: each bit selects or excludes a feature (Brookhouse et al., 2023, Imani et al., 2023). In healthcare workforce scheduling, a chromosome is a complete weekly schedule represented as a dictionary keyed by $i$1, whose values are lists of slot assignments over 48 half-hour intervals (Patel et al., 28 Aug 2025). In container allocation, each microservice is encoded by an allocation list whose length equals its scale level, so the same genetic structure controls elasticity and placement (Guerrero et al., 2024). In image registration, a chromosome contains six affine parameters: translation in $i$2 and $i$3, rotation, scale, and two shear terms (Chicotay et al., 2017). In accelerator optimization, each individual is a real-valued vector of cavity gradients (Rajput et al., 2024). In crystal structure prediction, the variables are free Wyckoff fractional coordinates in $i$4 (Yang et al., 2021).

Some encodings are explicitly hybrid. In decentralized exchange routing, the genotype is a variable-length set of execution paths together with continuous split ratios (Marfinetz, 24 Oct 2025). In industrial 3D pallet loading, GENPACK uses a layer-based chromosome

$i$5

where each layer stores assigned items, 2D coordinates, layer height, and base $i$6-level (Poolavaram et al., 16 Jan 2026). These designs show that MOO-GA frequently combines discrete topology decisions with continuous allocation variables.

Variation operators are equally specialized. Standard simulated binary crossover and polynomial mutation are used in MG-GPO and NSGA-III-based SNN fitting (Huang et al., 2019, Fitzgerald et al., 2021). CNN pruning instead uses microbial crossover, in which a lower-fitness winner infects a loser genewise, plus distinct mutation rates for convolutional and fully connected layers (Yang et al., 2019). Container allocation applies one-point crossover on allocation lists together with swap, shrink, and growth mutation, the last two directly changing replica counts (Guerrero et al., 2024). DEX routing uses edge-preserving crossover and mutations such as splice-add, splice-drop, edge-swap, pool-substitution, and ratio-perturb (Marfinetz, 24 Oct 2025). GENPACK mutates by moving tall items inward or downward, filling voids, compacting layers, and increasing side or bottom support (Poolavaram et al., 16 Jan 2026). The recurring pattern is that effective MOO-GA implementations rely on operators aligned with domain semantics rather than generic string perturbations.

4. Hybrid, surrogate-assisted, and model-driven variants

Many modern MOO-GAs augment genetic search with learned models, heuristics, or deterministic solvers. In “Evolutionary Multi-Objective Optimization Driven by Generative Adversarial Networks,” parent solutions are classified into real and fake samples at each generation, a GAN is trained on that split, and offspring are then sampled from the trained generator; the method is reported on 10 benchmark problems with up to 200 decision variables (He et al., 2019). Here the generative model effectively replaces part of the traditional crossover-mutation role.

MG-GPO is a surrogate-assisted evolutionary optimizer that keeps the generational MOEA structure but inserts Gaussian-process screening between candidate generation and expensive evaluation (Huang et al., 2019). For each objective, the method builds a GP surrogate and ranks trial solutions using the lower confidence bound

$i$7

then applies non-dominated sorting in surrogate space and evaluates only a selected subset on the true system. The paper reports substantially higher convergence speed and stability than NSGA-II, MOPSO, MMOPSO, and WOF-SMPSO on ZDT benchmarks (Huang et al., 2019).

Music harmony generation uses a different surrogate pattern. A first GA stage generates candidate pieces that are scored by experts and ordinary listeners; two Bi-LSTM regressors are then trained to predict those scores, and a second GA stage combines a grammar objective with the two learned evaluators in the scalar fitness (Majidi et al., 2021). The search remains evolutionary, but part of the objective function becomes learned rather than hand-coded.

Hybridization can also take the form of solver selection and fallback. In CoW Protocol routing, a production-grade NSGA-II engine is paired with adaptive instance profiling and a deterministic dual-decomposition optimizer; the deployed solution is chosen so as never to underperform the deterministic baseline in user surplus (Marfinetz, 24 Oct 2025). OTNSGA-II modifies a different part of the pipeline, replacing random initialization with orthogonal experimental design and using adaptive clustering pruning to eliminate similar or poor-quality individuals during survival selection (Yang et al., 2019). These examples indicate that current MOO-GA practice often treats the GA as one component in a broader optimization architecture rather than as a standalone search loop.

5. Diversity preservation and feasibility control

Maintaining spread and avoiding premature convergence remain central concerns. Standard crowding distance is widely used, but multiple papers introduce stronger diversity mechanisms. OTNSGA-II adds clustering-based pruning, removing individuals with similar features and poor non-dominated or crowding characteristics, or points judged far from the Pareto front according to intra-class aggregation (Yang et al., 2019). CMCrystalMOO goes further by making age itself an objective: all initial individuals start with age 1, survivors age by one generation, offspring inherit the maximum parental age, and 5% random individuals are injected each generation to keep old mediocre lineages from dominating the population (Yang et al., 2021). NSGA-III for SNN fitting uses normalization, association, and niche preservation after partial acceptance of the last front, reflecting a reference-direction approach to diversity (Fitzgerald et al., 2021).

Constraint handling is equally heterogeneous. In accelerator tuning, candidate RF-gradient vectors are evaluated under per-cavity bounds and a tight global energy constraint; infeasible solutions are assigned low fitness (Rajput et al., 2024). In fractional-order chaos control, controller gains are screened analytically against Matignon or Deng stability conditions before time-domain objectives are simulated, and unstable candidates receive large penalties (Pan et al., 2016). In image registration, a two-phase procedure first uses a feature-based Euclidean-distance objective for coarse alignment, then refines with Pareto-based simultaneous optimization of that feature distance and normalized cross correlation (Chicotay et al., 2017).

Industrial packing and scheduling illustrate a more procedural style of feasibility control. GENPACK combines structured representation, overlap penalties, support thresholds, post-processing compaction, fallback placement, and final validation that removes unsupported or overlapping items (Poolavaram et al., 16 Jan 2026). The healthcare scheduler relies on heavy penalties for hard constraint violations such as unavailability, split shifts, insufficient rest, and missing manager coverage, while still using NSGA-II as the survivor-selection engine (Patel et al., 28 Aug 2025). A plausible implication is that successful MOO-GAs often balance hard feasibility, soft penalties, and domain-specific repair rather than relying on a single universal constraint mechanism.

6. Application landscape

The supplied literature applies MOO-GA to a broad set of optimization domains. In routing and logistics, it appears in multiple-objective vehicle routing (0809.0416), healthcare workforce scheduling (Patel et al., 28 Aug 2025), industrial 3D pallet loading (Poolavaram et al., 16 Jan 2026), cloud container placement and elasticity management (Guerrero et al., 2024), and decentralized exchange order routing (Marfinetz, 24 Oct 2025). In scientific and engineering design, it is used for Cherenkov telescope array geometry (Souto et al., 2017), accelerator RF optimization (Rajput et al., 2024), crystal structure prediction (Yang et al., 2021), and active control of fractional-order chaotic financial systems (Pan et al., 2016).

In machine learning and data analysis, MOO-GA is used for CNN pruning under error, computation, and sparsity criteria (Yang et al., 2019), fair feature selection under accuracy and fairness metrics (Brookhouse et al., 2023), multi-view feature selection under classification error and subset size (Imani et al., 2023), and rigorous design of toxicity reference compound lists (Ohto et al., 11 May 2025). In image and sequence modeling, it appears in affine image registration (Chicotay et al., 2017) and polyphonic music generation that combines theory-based and learned human-preference objectives (Majidi et al., 2021). In computational neuroscience, NSGA-III is used to fit recurrent spiking-neural-network connectivity so as to match excitatory and inhibitory population firing rates, with sparse connectivity often emerging among the best solutions (Fitzgerald et al., 2021).

This application spread shows that MOO-GA is not confined to one ontology of “genetic” search. In some domains, it functions as a direct optimizer over physical or geometric variables; in others, it serves as a wrapper around classifiers, simulators, or learned evaluators. The shared feature is not the domain, but the use of population-based variation and selection to navigate explicit trade-offs among competing criteria.

7. Recurring limitations and debates

A persistent debate concerns what should count as “multi-objective.” Several papers explicitly note that a weighted aggregate scalar fitness is not a true Pareto MOEA, even when the underlying problem is multi-criteria (Yang et al., 2019, Majidi et al., 2021). The prey-predator hardness study reinforces this ambiguity by labeling an additive aggregation $i$8 as multi-objective while measuring hardness largely through convergence and non-convergence behavior (Ansari et al., 2014). This suggests that the term MOO-GA is used in both a strict Pareto sense and a looser multi-criteria sense.

Evaluation protocols are also inconsistent. MG-GPO reports hypervolume and inverted generational distance (Huang et al., 2019). Accelerator tuning evaluates Pareto fronts by hypervolume (Rajput et al., 2024). Image registration reports RMSE against manually selected points (Chicotay et al., 2017). Toxicity-list design measures downstream AUC, F1, accuracy, and additional statistical tests (Ohto et al., 11 May 2025). By contrast, some application papers provide mainly visual, practical, or demonstrative evidence rather than systematic MOEA benchmarking (0809.0416, Pan et al., 2016). Direct cross-paper comparison is therefore limited.

Reproducibility varies substantially. Several papers omit exact penalty weights, repair details, crossover or mutation specifics, or final decision rules over the returned trade-off set (Patel et al., 28 Aug 2025, Poolavaram et al., 16 Jan 2026). Others explicitly acknowledge ambiguities in formulas or pseudocode (Yang et al., 2019, Majidi et al., 2021). Computational burden remains another recurring issue: expensive simulation, repeated retraining, large combinatorial spaces, and tight feasibility regions all push practitioners toward surrogates, deterministic fallbacks, or strong heuristic seeding rather than pure GA search (He et al., 2019, Marfinetz, 24 Oct 2025). The literature therefore presents MOO-GA not as a settled canonical method, but as a flexible optimization framework whose effectiveness depends heavily on formulation choices, representation, and domain-specific hybridization.