Evolutionary Algorithms and Engineering Applications: A Comprehensive Survey of Classical Methods and Emerging Trends

1. Introduction

Evolutionary algorithms (EAs) constitute a central paradigm in computational intelligence and soft computing, providing population-based optimization frameworks inspired by the principles of natural selection and biological evolution [1,2]. Since their inception in the 1960s and 1970s, with foundational contributions from pioneers such as John Holland, Ingo Rechenberg, and Lawrence Fogel, EAs have evolved from conceptual biological metaphors into mature optimization methodologies with strong theoretical foundations and broad applicability. Representative classes include genetic algorithms (GAs) [3,4,5,6], genetic programming (GP) [7,8], differential evolution (DE) [9,10], evolution strategies (ES) [11,12], and evolutionary programming (EP) [13]. These approaches use iterative processes of variation, selection, and inheritance to effectively explore high-dimensional, multimodal, and rugged fitness landscapes where classical optimization methods often fail [14,15]. The enduring success of EAs arises from their ability to maintain population diversity, perform global search, and adaptively balance exploration and exploitation [16,17]. Their parallel evaluation of candidate solutions provides effectiveness against local optima, noise, and non-convexity, enabling widespread adoption in engineering design optimization [18], scheduling [19], image and signal processing, bioinformatics [20], and environmental modeling [21]. Furthermore, their flexible and modular nature facilitates hybridization with learning paradigms such as neural architecture search (NAS), reinforcement learning, and adaptive control [22]. In recent years, EA research has accelerated significantly, driven by advances in computation, availability of large-scale datasets, and integration with emerging fields such as quantum computing, AI, and big data analytics [23].

Although EAs have been widely reviewed, most prior surveys remain largely algorithm-centric, emphasizing taxonomies, operators, theory, and benchmark behavior. These works classify methods by their internal mechanics but provide limited insight into how different EA variants are adapted to the needs of specific engineering domains. With the rapid growth of domain-designed EA designs in areas such as energy systems, healthcare, and intelligent infrastructure, a synthesized view that links algorithmic choices to problem-driven requirements is increasingly needed.

This survey addresses that gap by offering a domain-centric synthesis of evolutionary algorithms. While covering classical EA families and modern hybrid approaches, the discussion is organized around major application domains where EAs have demonstrated significant impact. The goal is to help practitioners understand which EA variants suit particular problem characteristics, clarify how methodological developments translate to domain performance, and highlight trends emerging from real-world deployments. In addition, this survey consolidates theoretical foundations, methodological diversity, and cross-domain applications while highlighting open challenges such as scalability, convergence reliability, computational cost, and integration with large-scale and data-driven systems. The analysis underscores how EAs continue to evolve as essential tools for next-generation intelligent optimization.

2. Methodological Review Framework

This survey adopts a transparent and reproducible review protocol to ensure search completeness, consistent screening, and domain-relevant coverage.

2.1. Search Strategy

A comprehensive search was performed across IEEE Xplore, Scopus, Web of Science, ScienceDirect, and SpringerLink for the period 2015–2025. Core EA keywords included “evolutionary algorithm”, “genetic algorithm”, “differential evolution”, “evolution strategies”, “genetic programming”, “memetic algorithm”, “surrogate-assisted evolutionary algorithm”. To ensure methodological transparency and reproducibility, the complete Boolean search strings for all databases including explicit AND/OR structures and domain-specific filters are provided in the Supplementary Appendix (Appendix A). The initial retrieval yielded 2,183 records; after removing duplicates (412) and non-English works (76), 1,695 items remained for screening.

2.2. Screening and Selection Criteria

A total of 953 papers underwent full-text review. The refined inclusion and exclusion criteria are:

Inclusion Criteria:

• Peer-reviewed journal or conference papers (2015–2025);
• Evaluation on two or more benchmark functions or one real-world engineering case;
• Explicit description of EA components (representation, operators, selection);
• Sufficient experimental detail for replication (parameter settings, dataset/task description, and statistical reporting).

Exclusion Criteria:

• Pure swarm-intelligence studies lacking EA components;
• Works without replicable methodology (missing parameters, unclear benchmarks, incomplete metrics);
• Non-archival content (theses, non-reviewed preprints, posters, tutorials).

After applying these criteria, 312 publications formed the final corpus analyzed throughout the survey.

2.3. Classification and Taxonomy Approach

The selected works were organized using a dual-axis taxonomy comprising an algorithmic axis (GA, DE, ES/CMA-ES, GP, EP, hybrid EAs, and quantum-inspired EAs) and an application axis (healthcare, engineering design, energy systems, AI/ML, robotics, and smart cities). Figure 1 provides a unified visualization linking EA families to their dominant application domains.

Rationale Behind the Taxonomy: Different EA families possess distinct representations and search dynamics that naturally align with specific domain requirements. CMA-ES and related evolution strategies are favored in engineering and structural optimization due to their Gaussian sampling and covariance adaptation suited for high-dimensional, ill-conditioned landscapes. GAs with discrete encodings and recombination remain prevalent in routing, scheduling, and other combinatorial tasks. Differential evolution is widely applied in energy systems, control, and parameter estimation because of its simple vector-based mutation and noise resilience. Hybrid and memetic EAs dominate medical imaging and healthcare tasks, where domain heuristics, prior knowledge, and multi-objective trade-offs are critical. By integrating algorithmic and application axes, the taxonomy provides the domain-centric lens central to this survey, clarifying how methodological characteristics map onto real-world problem structures.

2.4. Study Overview

The final corpus spans one decade and covers classical, hybrid, surrogate-assisted, and quantum-inspired evolutionary algorithms.

3. Foundations of Evolutionary Algorithms

3.1. Core Mechanisms of Evolutionary Algorithms

Evolutionary algorithms are population-based stochastic optimizers driven by the principles of selection, recombination, and mutation. Although these mechanisms are standard and well documented in classical textbooks, we briefly summarize them here for completeness and refer readers to Table 1 for a compact overview of how each operator shapes exploration and exploitation. The generic evolutionary update can be written as

$$P_{t+1}=\mathcal{S}\left(\mathcal{M}\left(\mathcal{C}\left(P_t\right)\right)\right),$$

(1)

Here $\mathcal{S}$, $\mathcal{C}$, and $\mathcal{M}$ denote selection, crossover, and mutation, respectively. Algorithm 1 outlines the modular workflow. These components collectively enable EAs to maintain diversity, refine promising regions, and navigate rugged fitness landscapes.

Algorithm 1 Generic Evolutionary Algorithm Framework

Require:  Search space $\Omega$, fitness function $f:\Omega \to {\mathbb{R}}^k$, population size N, maximum generations $G_{max}$, crossover rate $p_c$, mutation rate $p_m$, selection operator $\mathcal{S}$, crossover operator $\mathcal{C}$, mutation operator $\mathcal{M}$, replacement operator $\mathcal{R}$ Ensure:  Best-so-far solution $x^{★}$ 1: $t\leftarrow 0$ 2: Initialize population $P^{\left(0\right)}={\left\{x_i^{\left(0\right)}\right\}}_{i=1}^N$ with $x_i^{\left(0\right)}\sim \mathcal{U}(\Omega )$ 3: Evaluate fitness $F^{\left(0\right)}={\left\{f\left(x_i^{\left(0\right)}\right)\right\}}_{i=1}^N$ 4: Set $x^{★}\leftarrow arg{min}_{x_i^{\left(0\right)}\in P^{\left(0\right)}}f\left(x_i^{\left(0\right)}\right)$ 5: repeat 6:     $t\leftarrow t+1$ 7:     Compute selection probabilities ${\pi }^{(t-1)}$ from $F^{(t-1)}$ 8:     $M^{(t-1)}\leftarrow \mathcal{S}\left(P^{(t-1)},{\pi }^{(t-1)}\right)$ 9:     Initialize offspring population $O^{\left(t\right)}\leftarrow \varnothing$ 10:     for $i=1$ to N step 2 do 11:         Select parents $x_p^{(t-1)},x_q^{(t-1)}\in M^{(t-1)}$ 12:         Draw $u\sim \mathcal{U}(0,1)$ 13:         if $u<p_c$ then 14:            $({\tilde{x}}_1^{\left(t\right)},{\tilde{x}}_2^{\left(t\right)})\leftarrow \mathcal{C}\left(x_p^{(t-1)},x_q^{(t-1)}\right)$ 15:         else 16:            ${\tilde{x}}_1^{\left(t\right)}\leftarrow x_p^{(t-1)}$;    ${\tilde{x}}_2^{\left(t\right)}\leftarrow x_q^{(t-1)}$ 17:         end if 18:         Draw $u_1,u_2\sim \mathcal{U}(0,1)$ 19:         if $u_1<p_m$ then 20:            ${\tilde{x}}_1^{\left(t\right)}\leftarrow \mathcal{M}\left({\tilde{x}}_1^{\left(t\right)}\right)$ 21:         end if 22:         if $u_2<p_m$ then 23:            ${\tilde{x}}_2^{\left(t\right)}\leftarrow \mathcal{M}\left({\tilde{x}}_2^{\left(t\right)}\right)$ 24:         end if 25:         $O^{\left(t\right)}\leftarrow O^{\left(t\right)}\cup \{{\tilde{x}}_1^{\left(t\right)},{\tilde{x}}_2^{\left(t\right)}\}$ 26:     end for 27:     Evaluate $F_O^{\left(t\right)}=\left\{f\left(x\right)\right|x\in O^{\left(t\right)}\}$ 28:     $P^{\left(t\right)}\leftarrow \mathcal{R}\left(P^{(t-1)},F^{(t-1)},O^{\left(t\right)},F_O^{\left(t\right)}\right)$ 29:     Update $x^{★}\leftarrow arg{min}_{x\in P^{\left(t\right)}\cup \left\{x^{★}\right\}}f\left(x\right)$ 30: until $t\ge G_{max}$ or a convergence/fitness criterion is satisfied 31: return  $x^{★}$

3.2. Theoretical Foundations and Their Practical Implications

A theoretical understanding of evolutionary algorithms provides insight into their expected convergence behavior, stability, and search efficiency. These results not only clarify why certain EA variants behave as they do but also guide practitioners in choosing suitable operators, mutation strengths, population sizes, and hybridization strategies. Below we synthesize key theoretical concepts and highlight their practical implications.

3.2.1. Markov-Chain Convergence

Many EAs can be modeled as homogeneous Markov chains $\left\{X_t\right\}$ on a finite state space $\mathcal{S}$. If mutation ensures positive probability of reaching any feasible solution, the chain becomes irreducible and aperiodic, implying a stationary distribution:

$$\underset{t\to \infty }{lim}Pr(X_t\in S^{★})=\pi \left(S^{★}\right)>0,$$

(2)

Here $S^{★}$ is the set of optimal populations [24,25].

Practical implication: algorithms with mutation operators that guarantee reachability (e.g., Gaussian noise, bit-flip mutation) avoid search stagnation and theoretically ensure long-term access to optimal regions.

3.2.2. Runtime Analysis and Drift Bounds

Runtime is often expressed as the hitting time to an optimal region:

$$\tau =min\{t\ge 0:X_t\in {\Omega }^{★}\}.$$

(3)

Drift analysis quantifies expected progress via a potential function ${\mathsf{\Phi }}_t$:

$$\mathbb{E}[{\mathsf{\Phi }}_t-{\mathsf{\Phi }}_{t+1}]\ge \eta \mathrm{or}\mathbb{E}[{\mathsf{\Phi }}_t-{\mathsf{\Phi }}_{t+1}]\ge \lambda {\mathsf{\Phi }}_t.$$

(4)

Additive drift leads to linear convergence; multiplicative drift yields logarithmic convergence.

Practical implication: algorithms that maintain diversity and proportional improvement achieve faster convergence on structured problems.

3.2.3. Landscape Properties: Discrete vs. Continuous

Discrete EAs rely on Hamming-neighborhood connectivity, whereas ES/CMA-ES use Gaussian sampling:

$$x_{t+1}\sim \mathcal{N}(m_t,{\sigma }_t^2{C}_t),$$

(5)

yielding linear contraction on smooth convex landscapes [2,26].

Practical implication: CMA-ES or continuous EAs are preferable for smooth, real-valued optimization, while bit-string EAs excel on combinatorial structures.

3.2.4. Self-Adaptation and Parameter Dynamics

Adaptive rules update parameters such as mutation strength:

$${\sigma }_{t+1}=\left\{\begin{array}{cc}\alpha {\sigma }_t,\hfill & \mathrm{success},\hfill \\ {\alpha }^{-1/4}{\sigma }_t,\hfill & \mathrm{otherwise},\hfill \end{array}\right.\alpha >1.$$

(6)

Practical implication: success-based or learning-driven parameter control is crucial on noisy or multi-modal landscapes where static settings fail.

3.2.5. Multiobjective Optimization Guarantees

Classical MOEAs satisfy polynomial-time approximation bounds:

$$\mathbb{E}\left[T_{\epsilon }\right]=\mathrm{poly}(n,1/\epsilon ),$$

(7)

with IGD and hypervolume converging under adequate diversity.

Practical implication: practitioners must ensure diversity preservation for reliable Pareto front approximation.

3.2.6. Constraint-Handling Theory

Penalty-based reformulations use:

$$F\left(x\right)=f\left(x\right)+\rho \sum _{i}max\{0,g_i\left(x\right)\},$$

(8)

while feasibility-based rules offer stronger guarantees under certain geometries.

Practical implication: constraint-handling should be aligned with the structure of feasible regions; feasibility-based rules often outperform penalties when constraints are tight or nonlinear.

3.2.7. No-Free-Lunch Limits

For uniform distributions over all objective functions, implying no universally superior optimizer.

$${\displaystyle \frac{1}{\left|\mathcal{F}\right|}}\sum _f{\mathrm{Perf}}_A\left(f\right)={\displaystyle \frac{1}{\left|\mathcal{F}\right|}}\sum _f{\mathrm{Perf}}_B\left(f\right),$$

(9)

Practical implication: this motivates the diversity of EA variants, hybridization strategies, and domain-specific adaptations discussed in subsequent sections. Designing operators, representations, and learning components to problem structure is essential for practical performance gains.

4. Types of Evolutionary Algorithms

Evolutionary algorithms constitute a broad family of population-based optimizers that differ in representation, variation operators, and selection strategies, leading to distinct performance characteristics across domains. Rather than cataloguing each method in isolation, this section adopts a domain-centric perspective: we summarize core algorithm families, highlight their strengths and limitations, and analyze where they are most suitable in engineering and scientific practice. The comparative taxonomies presented in Table 2 and Table 3 organize classical and advanced EA paradigms, respectively, linking algorithmic principles to their characteristic strengths, weaknesses, and ideal application contexts.

4.1. Genetic Algorithms (GAs)

GAs are population-based metaheuristics that evolve candidate solutions via selection, crossover, and mutation. Given a search space $\mathcal{X}\subseteq {\mathbb{R}}^d$ and objective function $f:\mathcal{X}\to \mathbb{R}$, the population ${\mathcal{P}}^{\left(g\right)}=\{x^{\left(1\right)},\dots ,x^{\left(N\right)}\}$ is iteratively evaluated, varied, and updated, as summarized in Algorithm 2. GAs support binary, real-valued, permutation, and graph encodings, making them broadly applicable in discrete and combinatorial domains.

Algorithm 2 Genetic Algorithm (GA)

Require:  Objective f, population size N, $p_c$, $p_m$, $G_{max}$ 1: Initialize population ${\mathcal{P}}^{\left(0\right)}$ 2: for $g=0$ to $G_{max}$ do 3:     Select mating pool ${\mathcal{M}}^{\left(g\right)}$ 4:     Apply crossover ($p_c$) and mutation ($p_m$) to form ${\mathcal{O}}^{\left(g\right)}$ 5:     Evaluate offspring and perform survivor selection 6: end for 7: return best solution found

GAs perform well when meaningful building blocks can be recombined, such as in scheduling, routing, feature selection, and structured design tasks [40,41,42,43,44,45,46,47]. However, their performance typically weakens on high-dimensional continuous landscapes where operators lack directional bias and where DE or CMA-ES are more effective. Recent advances include improved encodings and automated configuration [48,49], quantum-inspired and coevolutionary variants [39,50], and hybrid GA–ML or GA–RL frameworks [43,51,52,53,54] that enhance exploration and convergence efficiency.

Comparative Analysis and Domain Suitability

Compared with DE and CMA-ES, which operate directly on continuous vectors, GAs depend on encodings and schema-based recombination, making them most suitable for discrete, combinatorial, or structured search spaces. Relative to GP, GAs are less interpretable but simpler and more computationally efficient. Overall, GAs are preferred when problem structure enables meaningful recombination of partial solutions, while continuous optimization is better handled by DE or CMA-ES.

4.2. Evolution Strategies (ES, CMA-ES)

Evolution Strategies (ES) are derivative-free optimizers designed for continuous domains, with CMA-ES being the most influential modern variant. CMA-ES adapts a multivariate Gaussian sampling distribution, and updates its mean by weighted recombination of the best $\mu$ individuals:

$$x_k^{\left(g\right)}\sim \mathcal{N}\left(m^{\left(g\right)},{\left({\sigma }^{\left(g\right)}\right)}^2{C}^{\left(g\right)}\right),$$

(10)

$$m^{(g+1)}={\displaystyle \sum _{i=1}^{\mu }}{w}_i{x}_{i:\lambda }^{\left(g\right)}.$$

(11)

Step-size and covariance adaptation enable invariance to affine transformations and strong performance on non-separable, ill-conditioned landscapes. The core procedure is summarized in Algorithm 3. The $O\left(d^2\right)$ covariance update limits CMA-ES scalability in high dimensions, and reliable estimation requires large populations, giving high evaluation cost. Its Gaussian sampling also lacks a natural analogue in discrete or permutation spaces, making unmodified CMA-ES unsuitable for combinatorial problems. Consequently, lighter DE or PSO variants often outperform CMA-ES in low-dimensional or evaluation-constrained settings.

Algorithm 3 Covariance Matrix Adaptation ES (CMA-ES)

1: Initialize $m^{\left(0\right)},{\sigma }^{\left(0\right)},C^{\left(0\right)}=I$ 2: for $g=0$ to $G_{max}$ do 3:     Sample and evaluate $\lambda$ offspring 4:     Select top $\mu$ and update $m^{(g+1)},{\sigma }^{(g+1)},C^{(g+1)}$ 5: end for 6: return best solution

To reduce these issues, recent research explores reference vector–based ranking for multiobjective search [55], cooperative coevolution and landscape-aware grouping for large-scale optimization [56], noise-resilient population-size control [57], and parallel or chaotic variants [58,59]. Hybrid extensions broaden CMA-ES applications, including NAS [60], multimodal niching with PSO [61], microrobotics and trajectory planning [58,62], photonic device optimization [63,64], geophysical imaging [23,36], and wireless communication system design [65].

Comparative Analysis and Domain Suitability

Relative to GAs and DE, ES/CMA-ES provide principled step-size control and anisotropic search, making them highly effective on non-separable, ill-conditioned continuous problems. However, their quadratic update cost and reliance on continuous sampling make them less suitable for very high-dimensional or discrete tasks, where DE or GA variants tend to be more efficient. CMA-ES is therefore most appropriate when precise continuous optimization is required and sufficient evaluations are available to exploit local landscape geometry.

4.3. Differential Evolution (DE)

DE is a population-based optimizer for continuous spaces, defined by its geometry-driven mutation strategy. At generation g, each solution $x_i^{\left(g\right)}$ generates a donor vector

$$v_i^{\left(g\right)}=x_{r_1}^{\left(g\right)}+F\left(x_{r_2}^{\left(g\right)}-x_{r_3}^{\left(g\right)}\right),$$

(12)

followed by crossover with rate $C_r$ to produce a trial vector, and selection retains the fitter candidate (Algorithm 4). Its simplicity, directional search via population differences, and low computational cost make DE highly effective on multimodal continuous landscapes. DE’s main limitations stem from its reliance on continuous vector arithmetic making it poorly suited for categorical or combinatorial spaces without specialized encodings and its sensitivity to F and $C_r$, which can cause stagnation when diversity collapses. To address these issues, recent work proposes self-adaptive and collaborative DE variants for dynamic environments [66], Bayesian-assisted hybrid DE for antenna design [37], and fuzzy-logic or bidirectional DE for improved stability and search reliability [67,68,69]. Multiobjective and multitask extensions enhance performance on conflicting objectives [70,71,72], while surrogate-assisted DE accelerates expensive evaluations [73] at the cost of potential model bias. DE continues to see broad application in robotics [74], energy systems [75,76], ultrasonic sensing [77], plasma devices [78], power electronics [79], cybersecurity [80,81], scheduling [82], VANET security [83], and antenna optimization [84]. Fractional-order DE [85] further enhances exploration using historical dynamics, though at the expense of additional hyperparameters.

Algorithm 4 Differential Evolution (DE)

1: Initialize population ${\left\{x_i^{\left(0\right)}\right\}}_{i=1}^N$ 2: for $g=0$ to $G_{max}$ do 3:     for each $x_i^{\left(g\right)}$ do 4:         Mutation: $v_i^{\left(g\right)}=x_{r_1}^{\left(g\right)}+F(x_{r_2}^{\left(g\right)}-x_{r_3}^{\left(g\right)})$ 5:         Crossover: generate $u_i^{\left(g\right)}$ with rate $C_r$ 6:         Selection: $x_i^{(g+1)}=argmin\{f\left(x_i^{\left(g\right)}\right),f\left(u_i^{\left(g\right)}\right)\}$ 7:     end for 8: end for 9: return best solution

Comparative Analysis and Domain Suitability

Compared to GAs, which naturally suit binary and permutation encodings, DE excels in continuous domains due to its directional mutation. Relative to CMA-ES, DE is cheaper per generation and easier to tune, but lacks full covariance adaptation and is more prone to stagnation. Thus, DE is most suitable for medium-scale continuous optimization, whereas GA, CMA-ES, or domain-specific MOEAs are preferable for combinatorial, highly constrained, or strongly multiobjective tasks.

4.4. Genetic Programming (GP)

GP evolves symbolic structures trees, linear expressions, and grammars to produce interpretable models, making it well suited for symbolic regression and transparent decision-making. Semantics-aware variants [86,87] improve search efficiency and reduce bloat by guiding variation with functional meaning, though GP still faces high computational cost and uncontrolled structural growth. GP has been extended to multiobjective and large-scale tasks in classification, routing, and scheduling [88,89,90,91,92], where hierarchical or rule-based structures are advantageous. Its performance generally weakens in high-dimensional continuous spaces, where numerical optimizers such as DE or CMA-ES converge more reliably. Federated GP [93] reduce privacy constraints by evolving local models but introduces communication and heterogeneity challenges. In software engineering, GP supports program synthesis and metamorphic relation generation [94,95], offering interpretability benefits compared to LLM-based generators but limited by semantic complexity and long-range dependencies. Overall, GP is most effective when interpretability, symbolic reasoning, or structural flexibility is required, despite scalability and representation challenges (Table 4).

Comparative Analysis and Domain Suitability

Unlike GA, DE, or CMA-ES, which operate on fixed-length numeric encodings, GP searches over variable-length symbolic programs. This enables discovery of human-readable rules but increases computational overhead and risks bloat. For large-scale continuous optimization, DE or CMA-ES typically offer superior efficiency, whereas GP excels in domains requiring symbolic structure or interpretable models. GP therefore complements, rather than replaces, numeric evolutionary methods.

4.5. Multi-Objective and Many-Objective EAs

MOEAs extend single-objective EAs to optimize conflicting criteria such as accuracy, model size, and computational cost by generating Pareto optimal solutions [103,104]. MaOEAs further scale these ideas through dominance relaxation, indicator-based selection, and decomposition strategies to remain effective in high-dimensional objective spaces. MOEAs have become increasingly important in NAS, where accuracy, efficiency, and resource constraints must be optimized jointly. Representative methods include CNN-GA [105], AE-CNN [106], AE-CNN+E2EPP [107], MOEA-PS [108], NPENAS-NP [109], EEEA-Net-C [110], CGP-NAS [111], SMCSO [112], and SPNAS [113]. These frameworks integrate Pareto-based selection, probabilistic modeling, and self-adaptive mutation to co-optimize accuracy and computational efficiency. As summarized in Table 5, EA-based NAS methods achieve competitive CIFAR-10/100 performance while significantly reducing GPU-day requirements, with models such as MOEA-PS [108] and SMCSO [112] illustrating the growing importance of MOEA/MaOEA paradigms in resource-aware architecture discovery.

Comparative Analysis and Domain Suitability

MOEAs and MaOEAs provide principled mechanisms for discovering trade-offs without predefined weightings, making them well suited to engineering design, energy planning, and NAS. However, dominance-based methods degrade as objectives increase; hence MaOEAs rely on reference vectors or indicator-based criteria to retain selection pressure. In practice, MOEAs are preferred when the number of objectives is small and Pareto front interpretability is important, whereas MaOEAs are more appropriate for high-dimensional settings where user-defined preferences or performance indicators guide search.

4.6. Hybrid, Memetic, Surrogate-Assisted, Reinforcement Learning–Enhanced, and Quantum Evolutionary Algorithms

Advanced evolutionary paradigms integrate global EA search with learning, modeling, or local refinement to overcome limitations of classical EAs such as high evaluation cost, slow convergence, and weak exploitation.

4.6.1. Surrogate-Assisted Evolutionary Algorithms (SAEAs)

SAEAs couple EAs with predictive models to reduce evaluation cost in expensive optimization [118,119,120]. Surrogates include ANNs [121,122,123], BP networks [124], Gaussian Processes/Kriging [125,126,127,128], RBF models [126], SVR [129], and adaptive systems (ANFIS, BFNN) [130,131]. SAEAs follow iterative surrogate construction and EA-based search, supporting applications such as energy-efficient buildings [121,122,132], motors [124], aero-engines [131], antenna/ship design [125,126,127], manufacturing [123,128,130,133], and energy systems [129]. AutoSAEA [134] automates surrogate and infill selection. Despite efficiency gains, SAEAs face challenges of surrogate bias and drift in high-dimensional or noisy settings.

4.6.2. Memetic and Hybrid Evolutionary Algorithms

Memetic and hybrid EAs combine EA-based global exploration with heuristic or deterministic local search. Applications include drone–truck routing [135], vehicular offloading [136], satellite scheduling [137], berth–crane assignment [138], hybrid flow-shops [139], energy-aware job shops [140], setup-dependent scheduling [141], human–robot collaboration [142], and sustainable manufacturing [143]. They also support NAS [144], multimodal TSP [145], localization [146], polar codes [147], subpixel mapping [148], and facility layout [149]. While hybrids improve exploitation and convergence, they are sensitive to local-search design and domain heuristics.

4.6.3. Reinforcement Learning–Assisted Evolutionary Algorithms (RL-EAs)

RL-EAs adapt operators, parameters, or selection rules using learned policies [150]. Q-learning improves GA-based scheduling [151,152,153,154,155], while Dueling DQN aids satellite scheduling [156] and Actor–Critic supports production planning [157]. PPO stabilizes combinatorial search [158]. In DE, Q-learning and policy-gradient methods adjust mutation strategies [159,160,161]. RL also enhances ABC [162,163,164,165,166,167] and dynamic routing [168]. In MOEAs, RL selects objectives [169,170,171] or weights [172,173,174,175]. Further examples include PSO–RL [165,176], GP–RL [177], multitask MFEA–RL [178], and inverse-RL hybrids [179]. RL-EAs handle dynamic environments but incur training overhead and reward-sensitivity.

Comparative Analysis and Domain Suitability

SAEAs are most effective when evaluations are expensive (aerospace, simulation-driven design), but vulnerable to surrogate bias. Memetic/Hybrid EAs accelerate convergence where local refinement is meaningful (scheduling, manufacturing, NAS), though performance depends on problem-specific heuristics. RL-EAs excel in dynamic or heterogeneous environments (routing, resource allocation, multiagent coordination), but require extensive interactions and careful reward shaping. QEAs introduce quantum-inspired diversity and search operators, offering potential advantages in combinatorial problems but remain limited by hardware and simulator constraints.

Overall, these paradigms complement classical GA/DE/ES by addressing evaluation cost (SAEA), exploitation (memetic/hybrid), adaptivity (RL-EA), and diversity (QEA), enabling more efficient and domain-aware search in complex real-world applications.

Table 6. Recent applications and methodological advances of Surrogate-Assisted Evolutionary Algorithms (SAEAs) in engineering and optimization domains (2020–2025).

Field	Authors / Year	Surrogate Model(s)	Objectives	Evolutionary Framework
Application-oriented SAEAs
Building energy-efficient design	Bre et al. [121] 2020	ANN	Multi-objective	NSGA-II
	Gonçalves et al. [132] 2020	Adaptive surrogate	Multi-objective	NSGA-II
	Chegari et al. [122] 2021	ANN	Multi-objective	GA
Motor manufacturing	Li et al. [124] 2021	BP network	Multi-objective	MOPSO
Aero-engine compressor design	Baert et al. [131] 2020	BFNN	Multi-objective	Online SAEA
Antenna design	Zhang et al. [125] 2020	Gaussian Process	Single-objective	DE
	Yu et al. [126] 2020	Kriging, RBF, ANN	Single-objective	PSO, DE
Ship design	Wang et al. [127] 2021	Kriging	Single-objective	GA
Automobile design	Li et al. [130] 2022	ANFIS	Multi-objective	SSPEA
	Wang et al. [133] 2021	RSM, Kriging	Multi-objective	MOGA
	Su et al. [123] 2021	ANN	Multi-objective	NSGA-II
Wing optimization	Wansaseub et al. [128] 2020	Kriging	Multi-objective	Latin Hypercube + DE
Energy and power	Ma et al. [129] 2021	SVR	Multi-objective	NSGA-II
Methodological advances in SAEAs
Cross-domain (expensive many-objective optimization)	Zhai et al. [133] 2023	Global + Local Kriging (composite surrogate)	Many-objective	Composite SAEA with filling sampling criterion

Table 6. Recent applications and methodological advances of Surrogate-Assisted Evolutionary Algorithms (SAEAs) in engineering and optimization domains (2020–2025).

Field	Authors / Year	Surrogate Model(s)	Objectives	Evolutionary Framework
Application-oriented SAEAs
Building energy-efficient design	Bre et al. [121] 2020	ANN	Multi-objective	NSGA-II
	Gonçalves et al. [132] 2020	Adaptive surrogate	Multi-objective	NSGA-II
	Chegari et al. [122] 2021	ANN	Multi-objective	GA
Motor manufacturing	Li et al. [124] 2021	BP network	Multi-objective	MOPSO
Aero-engine compressor design	Baert et al. [131] 2020	BFNN	Multi-objective	Online SAEA
Antenna design	Zhang et al. [125] 2020	Gaussian Process	Single-objective	DE
	Yu et al. [126] 2020	Kriging, RBF, ANN	Single-objective	PSO, DE
Ship design	Wang et al. [127] 2021	Kriging	Single-objective	GA
Automobile design	Li et al. [130] 2022	ANFIS	Multi-objective	SSPEA
	Wang et al. [133] 2021	RSM, Kriging	Multi-objective	MOGA
	Su et al. [123] 2021	ANN	Multi-objective	NSGA-II
Wing optimization	Wansaseub et al. [128] 2020	Kriging	Multi-objective	Latin Hypercube + DE
Energy and power	Ma et al. [129] 2021	SVR	Multi-objective	NSGA-II
Methodological advances in SAEAs
Cross-domain (expensive many-objective optimization)	Zhai et al. [133] 2023	Global + Local Kriging (composite surrogate)	Many-objective	Composite SAEA with filling sampling criterion

Table 7. RL-Assisted Evolutionary Algorithms by EA Type.

EA Type	Problem Domain	RL Role
Genetic Algorithm (GA)
Q-learning [151,152,153,154,155]	Scheduling, team formation	Operator control, task allocation
Dueling DQN [156]	Satellite scheduling	Dual-state evaluation
Actor–Critic [157]	Steel scheduling	Adaptive trade-off learning
PPO [158]	TSP, VRP, bin packing	Stable policy optimization
Differential Evolution (DE)
Variational PG [159]	Continuous SOP	Stochastic policy mutation
Policy Gradient [160]	Parameter tuning	Mutation rate control
Q-learning [161]	Trajectory design	Reward-based selection
Artificial Bee Colony (ABC)
Q-learning [162,163,165,166,167,180]	Flow-shop, traffic	Sequence, allocation rules
DQN [168]	Vehicle routing	Route exploration
MOEAs (MA, MOEA/D, NSGA-II/III)
Q-learning [172,173,175,181]	Job/flow-shop	Pareto weight tuning
DQN [169,170,171]	Energy, cloud	Objective adaptation
Hyper-Heuristic / Ensemble
Q-learning [151,182,183]	Energy, routing	Heuristic selection
DDQN / Double Q [184]	Packing, scheduling	Reward stability
Other EAs
PSO (Q-learning) [165,176]	Assembly, SOP	Velocity tuning
GP (Q-learning) [177]	Team formation	Task matching
MFEA (Q-learning) [178]	Multitask	Task transfer learning
MFO (Inverse RL) [179]	SOP	Expert reward imitation

Table 7. RL-Assisted Evolutionary Algorithms by EA Type.

EA Type	Problem Domain	RL Role
Genetic Algorithm (GA)
Q-learning [151,152,153,154,155]	Scheduling, team formation	Operator control, task allocation
Dueling DQN [156]	Satellite scheduling	Dual-state evaluation
Actor–Critic [157]	Steel scheduling	Adaptive trade-off learning
PPO [158]	TSP, VRP, bin packing	Stable policy optimization
Differential Evolution (DE)
Variational PG [159]	Continuous SOP	Stochastic policy mutation
Policy Gradient [160]	Parameter tuning	Mutation rate control
Q-learning [161]	Trajectory design	Reward-based selection
Artificial Bee Colony (ABC)
Q-learning [162,163,165,166,167,180]	Flow-shop, traffic	Sequence, allocation rules
DQN [168]	Vehicle routing	Route exploration
MOEAs (MA, MOEA/D, NSGA-II/III)
Q-learning [172,173,175,181]	Job/flow-shop	Pareto weight tuning
DQN [169,170,171]	Energy, cloud	Objective adaptation
Hyper-Heuristic / Ensemble
Q-learning [151,182,183]	Energy, routing	Heuristic selection
DDQN / Double Q [184]	Packing, scheduling	Reward stability
Other EAs
PSO (Q-learning) [165,176]	Assembly, SOP	Velocity tuning
GP (Q-learning) [177]	Team formation	Task matching
MFEA (Q-learning) [178]	Multitask	Task transfer learning
MFO (Inverse RL) [179]	SOP	Expert reward imitation

5. Applications of Evolutionary Algorithms

Evolutionary algorithms have been successfully applied across a wide range of scientific and engineering domains due to their global search capability, flexibility in handling diverse problem structures, and effectiveness in multi-objective and constraint-aware optimization. This section provides a domain-balanced overview of key application areas, focusing on healthcare, energy systems, robotics, and smart cities. Table 8 summarizes representative works across diverse fields.

5.1. Healthcare and Biomedical Applications

EAs play a central role in complex clinical optimization tasks such as radiation therapy planning, where multi-objective EAs optimize beam angles, dose distributions, and organ-at-risk protection [185,186,187,188]. Advanced MOEAs incorporate clinical priors and pattern-mining strategies to improve treatment diversity and stability [189]. In genomics, GAs and DE aid in disease-gene identification and feature selection from high-dimensional datasets [190,191]. In medical image segmentation, evolutionary NAS yields efficient architectures for few-shot and domain-shift scenarios [192,193,194,195].

Limitations and Open Challenges: EAs face obstacles including high computational cost for 3D medical imaging, difficulty integrating privacy regulations (HIPAA/GDPR), and limited interpretability in clinical pipelines. Scaling evolutionary NAS to large medical datasets also remains challenging.

5.2. Energy Systems and Smart Grid Optimization

Energy systems require effective optimization under uncertainty, renewable variability, and stringent operational constraints. EAs have been widely used for unit commitment, economic dispatch, and renewable energy planning [196,197]. Differential evolution and MOEAs are widely used for optimizing photovoltaic systems, microgrids, and wind-farm layouts due to their resilience to noisy, nonlinear objectives. Hybrid surrogate-assisted EAs further accelerate simulations for building energy design [121,122].

Limitations and Open Challenges: Dynamic, real-time grid operation demands faster convergence than classical EAs typically provide. Large-scale grid models challenge scalability, and surrogate drift may affect reliability. Future directions include hardware-aware EAs, online adaptation, and federated evolutionary control for geographically distributed systems.

5.3. Robotics, Control, and Autonomous Systems

Robotics applications use EAs for control tuning, trajectory planning, swarm coordination, and behavior synthesis. CMA-ES and DE excel in continuous control tasks such as manipulation, aerial robotics, and microrobotics [62,63]. Genetic programming and hybrid EAs evolve interpretable controllers and task policies, while evolutionary RL approaches enable adaptation in dynamic environments [151,152,158].

Limitations and Open Challenges: Real-time constraints restrict population sizes and evaluation budgets. Simulation-to-reality gaps reduce transferability, and safety-aware optimization remains underexplored. High-dimensional robotic systems necessitate more scalable, sample-efficient evolutionary controllers.

5.4. Smart Cities, Transportation, and Logistics

EAs support infrastructure planning, transportation scheduling, routing, and multi-criteria decision-making in smart cities. Applications include lane-reservation systems [198], traffic routing [168], supply-chain optimization [199], and cyber-physical coordination in IoT environments [200]. Multi-objective EAs enable trade-offs between efficiency, emissions, and congestion key factors in sustainable urban mobility.

Limitations and Open Challenges: Smart-city systems introduce dynamic, multi-agent environments where objectives evolve over time. Real-time EAs must operate under communication constraints and partial observability. Federated EAs, RL-EA hybrids, and decentralized optimization frameworks are promising future directions.

5.5. Artificial Intelligence and Machine Learning

EAs are increasingly integrated into AI pipelines for feature selection, hyperparameter tuning, and neural architecture search [201,202]. Multi-modal learning setups employ evolutionary multi-objective optimization for sensor fusion and model selection [203,204]. In secure IoT systems, post-quantum EA–blockchain frameworks enhance resilience and privacy [200]. Evolutionary NAS (ENAS) designs efficient architectures, while hybrid EA–RL and EA–gradient approaches balance accuracy and computational cost [205,206]. As shown in Figure 2, ENAS methods are generally categorized into three major evolutionary paradigms: EA, Swarm Intelligence, and Hybrid Search Strategies.

Limitations and Open Challenges: Evolutionary NAS remains computationally expensive. Hybrid EA LLM systems introduce interpretability and consistency challenges. Distributed AI introduces synchronization and latency issues that complicate evolutionary search.

6. Challenges and Problems in Evolutionary Algorithms

EAs face persistent challenges that limit their scalability, effectiveness, and deployment in modern large-scale systems. Figure 3 summarizes the major issues and commonly adopted mitigation strategies.

6.1. Scalability

High-dimensional problems expand the search space exponentially, requiring larger populations and stronger diversity preservation. CMA-ES suffers from $O\left(d^2\right)$ covariance updates, while L-SHADE and related DE variants scale better through self-adaptation. Modern settings such as federated or privacy-constrained optimization further strain scalability since subpopulations operate with limited coordination.

Promising Reduction Approaches: Cooperative and decomposition-based EAs offer dimensionality reduction but rely on accurate variable grouping; surrogate-assisted models reduce evaluation cost but introduce approximation bias; parallel/distributed EAs provide speedups but increase communication overhead; hardware-aware GPU-accelerated implementations help, though irregular operators limit full parallelism.

6.2. Balancing Exploration and Exploitation

Over-exploration slows convergence, whereas excessive exploitation causes premature stagnation. Static mutation or crossover rates struggle in dynamic or multimodal landscapes. RL-assisted EAs and island models provide adaptive control but require extra computation and careful design.

Promising Reduction Approaches: Self-adaptive operators, niching and clustering for diversity preservation, hybrid global local search, and RL-based operator control enhance balance, though each introduces risks such as oscillation, slower convergence, or additional training overhead.

6.3. Parameter Sensitivity and Self-Adaptation

EA performance depends heavily on mutation rates, crossover probabilities, population size, and selection pressure. Sensitivity becomes more severe in multiobjective and large-scale optimization. Self-adaptive and AutoML-driven tuning [227] reduce manual effort but add extra meta-parameters and computational burden.

Promising Reduction Approaches: Co-evolving parameters, meta-EAs, and learning-guided control mechanisms help improve parameter adaptation, but their performance depends heavily on population size and stable reward signals.

6.4. Computational Cost and Efficiency

Population-based evaluation makes EAs expensive for simulation-based design, NAS, robotics, and multi-physics tasks. Even with GPU parallelism, evaluating thousands of candidates per generation remains a bottleneck.

Promising Reduction Approaches: Surrogate-assisted EAs reduce evaluation cost but risk search bias; weight-sharing and pruning accelerate NAS but may distort fitness ranking; quantum-inspired and probabilistic EAs improve global exploration but remain small-scale; hardware-aware EAs show promise but depend on platform-specific tuning.

6.5. Benchmarking, Reproducibility, and Comparison

Reproducibility is hindered by heterogeneous implementations, inconsistent reporting, and sensitivity to random seeds or hardware differences. Distributed and federated optimization further complicate reproducibility due to communication delays and data shifts.

Promising Reduction Approaches: Standardized benchmark suites (e.g., COCO, LSGO, NAS-Bench), open-source repositories, statistical testing with effect sizes, and emerging federated benchmarking infrastructures improve transparency but require community-wide adoption.

7. Future Research Directions and Trends

Evolutionary algorithms are entering a transformative phase driven by the dual need for scalability and intelligence. Their future trajectory is shaped by two primary drivers: (1) Hybridization with complementary computational paradigms, and (2) Integration with emerging technologies such as quantum computing, LLMs, and federated systems. These drivers are supported by three foundational pillars that must co-evolve: (1) Theoretical grounding, (2) Domain-scale real-world deployment, and (3) Interpretability and trust. Together, these elements form a convergent roadmap for next-generation evolutionary computation.

7.1. Hybrid Evolutionary Algorithms as the Central Driver

Hybrid evolutionary algorithms (HEAs) will remain the cornerstone of future EA research. The ability to combine population-based exploration with gradient-based refinement, swarm heuristics, reinforcement learning, or local search enables the creation of adaptive and problem-aware optimizers. A key future direction involves dynamic operator orchestration, where hybrid components are selected or weighted based on real-time landscape feedback. This includes RL-driven operator scheduling, adaptive switching between exploration modes, and learning-guided pruning of ineffective search branches. Such hybridization is not only algorithmic but also architectural: future EAs will behave as self-evolving systems that restructure their own operators and representations during the optimization process.

7.2. Integration with Emerging Technologies: The Expanding Frontier

Hybridization intersects directly with technological advances, creating several emerging pathways for EA development.

7.2.1. Quantum Computing and Quantum-inspired EAs

Quantum-inspired EAs offer richer search distributions, parallel exploration, and theoretical speedups. Beyond classical optimization, methods like QDistEvol [228] will increasingly target problems native to quantum devices, including quantum error correction, quantum control, and circuit layout. A specific research direction is the development of quantum–classical co-evolution, where quantum subroutines evaluate fitness or generate offspring, while classical EAs guide global population dynamics.

7.2.2. Federated and Edge Evolutionary Computation

Federated and distributed environments demand optimization under privacy, communication, and heterogeneity constraints. Federated surrogate-assisted EAs [35,229] and evolutionary game-theoretic models [230,231,232] provide early foundations, but substantial gaps remain. Future directions include:

• Theoretical models of communication-efficient evolution under limited bandwidth,
• Incentive-compatible EAs for multi-agent and multi-owner data settings,
• Distributed multi-objective evolution with partial or inconsistent objective visibility.

These directions are essential for deploying EAs in smart cities, IoT ecosystems, and cyber-physical infrastructure.

7.3. Large Language Models (LLMs) as Evolutionary Meta-Controllers

LLMs are emerging as powerful agents capable of automating EA operator design, mutation generation, and adaptation logic. Frameworks such as LLaMEA [233,234] demonstrate that LLMs can synthesize new heuristics, self-correct search failures, and propose operator modifications using linguistic feedback. Figure 4 illustrates this co-evolution paradigm. Future work includes:

• LLM-driven operator innovation: generating new mutation/crossover families conditioned on landscape descriptors.
• Self-reflective evolution: EAs provide search logs, and LLMs respond with operator adjustments.
• Hybrid symbolic–numeric search: LLMs evolve symbolic rules while EAs refine numeric parameters.

This synergy could create autonomous evolutionary frameworks capable of reconfiguring themselves without human intervention.

7.4. Real-world Applications and Societal Impact

EAs are increasingly positioned to address global challenges across climate science, energy sustainability, personalized healthcare, and resilient urban infrastructure. Unlike conventional optimizers, their multi-objective nature enables them to balance conflicting goals such as energy efficiency versus emissions, accuracy versus interpretability, or cost versus resilience. Future research should target:

• Decarbonization and renewable energy systems: optimizing multi-scale models for smart grids, energy storage, and demand forecasting.
• Climate modeling and environmental resilience: integrating uncertainty-aware EAs for long-term climate scenario simulations.
• Precision medicine: evolving interpretable and privacy-preserving diagnostic pipelines.
• Real-time robotics and autonomous systems: building fast, hardware-aware EAs for dynamic control.

These directions align EAs with major societal missions such as sustainability, healthcare equity, and global risk mitigation.

7.5. Theoretical Foundations to Support Next-Generation EAs

Despite strong empirical success, theoretical understanding lags behind algorithmic innovation. Future theory must move beyond asymptotic convergence and provide actionable insights for modern algorithm families. Specific future directions include:

• Runtime and stability analysis of RL-EAs: identifying when learned adaptation policies outperform static operators.
• Theoretical models for federated evolution: quantifying how communication delays, heterogeneous data, and partial participation shape convergence.
• Landscape-aware operator theory: linking mutation/crossover dynamics to curvature, modality, or gradient surrogate information.
• Complexity bounds for hybrid and quantum-inspired EAs: developing computable performance guarantees for multi-layered or quantum-driven search.

These theoretical developments are key to improving the predictability, reliability, and safety of hybrid EA systems.

7.6. Enhanced Interpretability and Trustworthiness

As EAs move into medical diagnosis, financial decision-making, defense, and public infrastructure, interpretability becomes essential. Emerging work on visualizing evolutionary trajectories, tracing operator contributions, and extracting symbolic approximations of evolved solutions marks early progress. Future directions include:

• Causal interpretability: understanding which operators or genetic components drive solution improvements.
• Human-in-the-loop evolution: enabling experts to guide search trajectories interactively.
• Explainable multi-objective trade-offs: visualizing preference changes and Pareto dynamics in real time.

These tools will increase trust and accountability in domains where evolution-based decisions must be auditable.

7.7. Summary of Findings

EAs are evolving into intelligent optimization frameworks that combine evolutionary search with learning, distributed computation, and quantum-inspired mechanisms. Hybridization and integration with emerging technologies form the core research engine, supported by advances in theory, interpretability, and domain-focused deployment. These developments strengthen the role of EAs as scalable, transparent, and societally impactful optimizers.

7.8. Recommendations for Practitioners and Researchers

Practitioners should exploit domain knowledge in fitness modeling, adopt hybrid and automated tuning strategies, and prefer interpretable or resource-aware designs for real-world deployment. Researchers should pursue deeper theory for adaptive and learning-driven EAs, explore quantum and federated evolutionary paradigms, and prioritize transparent, reproducible, and societally aligned EA methodologies.

8. Conclusion

Evolutionary algorithms have matured into a diverse and powerful class of optimization methods, but their effectiveness is fundamentally shaped by the interaction between algorithmic design and domain characteristics. This survey adopted a domain-centric perspective, synthesizing EA foundations, algorithm families, and applications across engineering, energy, healthcare, transportation, and AI. By linking representation choices, variation operators, convergence behavior, and hybridization potential to domain requirements, the survey offers a structured guide for selecting suitable EA approaches. Several overarching insights emerge. First, no single EA is universally superior performance depends on fitness landscape structure, constraints, and evaluation cost. Second, hybridization has become the dominant trend, integrating evolutionary search with machine learning, surrogate modeling, reinforcement learning, and quantum-inspired mechanisms to improve scalability and effectiveness. Third, the increasing use of EAs in safety-critical and data-sensitive environments highlights the need for stronger theoretical guarantees, transparency, and interpretability. Looking forward, EAs are poised to evolve from standalone optimizers to essential components of next-generation intelligent systems. Their strengths in handling uncertainty, multiobjective trade-offs, and distributed information make them well-suited for emerging challenges in sustainable energy, climate resilience, personalized medicine, autonomous systems, and large-scale AI. As computational infrastructures and hybrid paradigms advance, evolutionary computation will continue to play a central role in optimization-driven intelligence for complex and dynamic real-world systems.

Appendix A.1. IEEE Xplore

(("evolutionary algorithm" OR "genetic algorithm" OR "genetic programming" OR"differential evolution" OR "evolution strategy" OR "memetic algorithm" OR"surrogate-assisted evolutionary algorithm") AND ("engineering" OR "optimization" OR"design" OR "medical" OR "energy" OR "robotics" OR "smart systems")) AND (Publication Year: 2015–2025)

Appendix A.2. Scopus

TITLE-ABS-KEY("evolutionary algorithm" OR "genetic algorithm" OR "genetic programming" OR "differential evolution" OR "evolution strategy" OR "memetic algorithm" OR"surrogate-assisted evolutionary algorithm" ) AND TITLE-ABS-KEY(engineering OR medical OR energy OR robotics OR design) AND (PUBYEAR > 2014)

Appendix A.3. Web of Science

TS = ("evolutionary algorithm" OR "genetic algorithm" OR "genetic programming" OR"differential evolution" OR "evolution strategy" OR "memetic algorithm") AND TS =("optimization" OR "engineering" OR "medical imaging" OR "energy system" OR "robotics" OR "smart city")

Appendix A.4. SpringerLink / ScienceDirect

("evolutionary algorithm" OR

"genetic algorithm" OR

"differential evolution" OR

"evolution strategy" OR

"genetic programming" OR

"memetic algorithm")

AND

("engineering" OR "medical" OR "energy" OR "robotics" OR "control")

AND

(year:[2015 TO 2025])