Data-Driven Differential Evolution

• Data-Driven Differential Evolution is a family of optimization methods that incorporates population geometry, historical data, and learned models to steer the search process.
• These methods enhance performance using techniques like clustering-guided mutation, reinforcement learning for operator selection, and surrogate models for complex industrial applications.
• Empirical results indicate that while enhancements in convergence and robustness are evident, performance gains remain problem-dependent and balanced against computational overhead.

Data-Driven Differential Evolution denotes a family of Differential Evolution (DE) methods in which search is guided by information extracted from the evolving population, from historical optimizer behavior, or from learned surrogate models, rather than relying only on ordinary random tuple sampling and fixed hand-designed operators. In the literature, the term covers several distinct regimes: population-geometry-driven DE, clustering-guided mutation, nearest-better neighborhood guidance, reinforcement-learning-based operator control, surrogate-assisted industrial DE, and meta-level evolution of DE configurations (Tomczak et al., 2020, Mousavirad et al., 2021, Dong et al., 2018, Sharma et al., 2019, Garate-Perez et al., 15 Jul 2025, Chen et al., 13 Feb 2025). The phrase is therefore best understood as a spectrum rather than a single algorithmic template.

1. Scope and conceptual usage

Within the DE literature, “data-driven” is used with different strengths. In a broad sense, a method is data-driven if it exploits search information beyond the immediate standard mutation formula. Under that definition, methods that use current-population geometry, cluster statistics, fitness-distance relations, run-state feedback, or offline learned policies all belong to the category. In a narrower sense, the label is reserved for explicit learned models, historical archives, surrogate models, or online adaptation from search traces. This difference is made explicit by work that describes reversible-transform DE as “population-data-driven DE” rather than model-based or surrogate-assisted DE (Tomczak et al., 2020).

A compact way to organize the literature is to distinguish where the “data” come from and how they enter the algorithm.

Regime	Main information source	Representative papers
Population-data-driven	Current population geometry or statistics	(Tomczak et al., 2020, Mousavirad et al., 2021, Dong et al., 2018)
Feedback-driven	Run-time progress and diversity signals	(Li et al., 2020)
Learned-control DE	Offline data from many DE runs	(Sharma et al., 2019, Chen et al., 13 Feb 2025)
Surrogate-assisted DE	Historical process data and learned objective/constraint models	(Garate-Perez et al., 15 Jul 2025)

This classification suggests that the unifying theme is not a single operator but a shift from purely unstructured stochastic search toward search informed by accumulated evidence. The evidence may be geometric, statistical, behavioral, or predictive.

2. Population-guided mutation and recombination

The most direct form of data-driven DE uses the current population itself as the data source. In standard DE/rand/1/bin, a mutant is generated as

$$\mathbf{y}=\mathbf{x}_i+F(\mathbf{x}_j-\mathbf{x}_k),$$

followed by binomial crossover

$$\mathbf{v}=\mathbf{m}\odot \mathbf{y}+(1-\mathbf{m})\odot \mathbf{x}_i,$$

with crossover probability $p=0.9$. Population-data-driven variants retain this framework but replace purely random operator use with transformations informed by observed population structure (Tomczak et al., 2020).

A notable example is Differential Evolution with Reversible Linear Transformations, whose main variant, RevDE, treats a sampled triplet as a local geometric configuration and generates three correlated offspring before any fitness evaluation:

$$\begin{aligned} \mathbf{y}_1 &= \mathbf{x}_i + F(\mathbf{x}_j - \mathbf{x}_k),\ \mathbf{y}_2 &= \mathbf{x}_j + F(\mathbf{x}_k - \mathbf{y}_1),\ \mathbf{y}_3 &= \mathbf{x}_k + F(\mathbf{y}_1 - \mathbf{y}_2). \end{aligned}$$

The transform has linear form $\mathbf{Y}=\mathbf{R}\mathbf{X}$ with $\det(\mathbf{R})=1$, so it is reversible for all $F$. The paper’s interpretation is that the method enriches the candidate set “for free” in mutation-source information, even though the enlarged offspring population still requires evaluation. This places RevDE squarely in the population-geometry-driven branch of data-driven DE (Tomczak et al., 2020).

Clu-DE uses a different population-derived signal. Each generation, it applies $k$-means clustering to the current population, scores clusters by mean objective function value, chooses the cluster with the lowest mean OFV as the winner cluster, and then uses the best individual inside that cluster as the base vector in a new mutation operator:

$\overrightarrow{v^{clu}_{i}}=\overrightarrow{winner}+F(\overrightarrow{x_{r1}-\overrightarrow{x_{r2}}).$

The method is therefore region-aware rather than purely random-base-driven. Its mutation remains stochastic through the difference term, but its base vector is selected from a subpopulation judged promising by cluster-level evidence (Mousavirad et al., 2021).

NbDE injects a more local population topology. For each current individual $x_i$, it identifies $$\mathbf{v}=\mathbf{m}\odot \mathbf{y}+(1-\mathbf{m})\odot \mathbf{x}_i,$$0, the nearest individual with better fitness, and uses that relation in mutation:

$$\mathbf{v}=\mathbf{m}\odot \mathbf{y}+(1-\mathbf{m})\odot \mathbf{x}_i,$$1

Here the relevant data are neither cluster summaries nor reversible transforms, but fitness ordering and Euclidean neighborhood structure among individuals. The resulting search direction is individualized: each target follows its own nearest superior neighbor rather than a common global best (Dong et al., 2018).

These methods show that “data-driven” need not imply a surrogate model. It can instead mean that existing individuals are treated as structured observations of search state, from which extra directions, promising regions, or local guides are inferred.

3. Learned operator control and meta-level configuration

A stronger interpretation of data-driven DE uses data collected across many optimization runs to learn or evolve the control logic of DE itself. In this regime, the algorithm is no longer only population-informed within a run; it is trained or configured from accumulated historical search behavior.

DE-DDQN is a clear example. It formulates mutation-strategy selection as a reinforcement-learning problem and uses Double Deep Q-Learning to choose among four mutation strategies—rand/1, rand/2, rand-to-best/2, and curr-to-rand/1—for each parent at each generation. The state is a 99-dimensional feature vector containing current fitness/progress features, parent-relative geometry and fitness relations, operator-history features over recent generations, and window-based historical features. Training is offline: the controller is trained on 16 CEC2005 functions in dimensions 10 and 30, then deployed without further learning on 5 different test functions. Among three reward definitions, the best-performing version is DE-DDQN2, using

$$\mathbf{v}=\mathbf{m}\odot \mathbf{y}+(1-\mathbf{m})\odot \mathbf{x}_i,$$2

The reported result is that DE-DDQN outperforms the non-adaptive methods for all functions in the test set, while remaining comparable with the two winners of the CEC2005 competition (Sharma et al., 2019).

MetaDE moves the same idea to a meta-optimization level. An outer DE evolves the configuration of an inner, parameterized DE family. Each outer individual encodes

$$\mathbf{v}=\mathbf{m}\odot \mathbf{y}+(1-\mathbf{m})\odot \mathbf{x}_i,$$3

where $$\mathbf{v}=\mathbf{m}\odot \mathbf{y}+(1-\mathbf{m})\odot \mathbf{x}_i,$$4 and $$\mathbf{v}=\mathbf{m}\odot \mathbf{y}+(1-\mathbf{m})\odot \mathbf{x}_i,$$5 specify left and right base vectors, $$\mathbf{v}=\mathbf{m}\odot \mathbf{y}+(1-\mathbf{m})\odot \mathbf{x}_i,$$6 is the number of differential vectors, and $$\mathbf{v}=\mathbf{m}\odot \mathbf{y}+(1-\mathbf{m})\odot \mathbf{x}_i,$$7 is the crossover scheme. The generalized PDE mutation is

$$\mathbf{v}=\mathbf{m}\odot \mathbf{y}+(1-\mathbf{m})\odot \mathbf{x}_i,$$8

yielding $$\mathbf{v}=\mathbf{m}\odot \mathbf{y}+(1-\mathbf{m})\odot \mathbf{x}_i,$$9 strategy combinations before considering continuous $p=0.9$0 and $p=0.9$1. MetaDE evaluates each candidate configuration by instantiating an executor DE on the target problem, so the outer search is driven directly by observed solver performance rather than by a handcrafted update rule. The framework is explicitly GPU-oriented and, under a 60-second budget on CEC2022, reaches about $p=0.9$2 function evaluations, whereas compared DE variants are around $p=0.9$3 (Chen et al., 13 Feb 2025).

Both methods are more strongly data-driven than purely population-geometric approaches. DE-DDQN learns a state-to-strategy mapping from offline trajectory data, whereas MetaDE evolves DE configurations from observed inner-run performance. In each case, DE control is derived from empirical evidence rather than specified a priori.

4. Surrogate-assisted and industrial data-driven DE

A still stronger form of data-driven DE appears when the objective and part of the constraint structure are learned from historical process data. In that setting, DE is not merely guided by search statistics; the evaluation function itself is data-derived.

The tire-industry extrusion study proposes “Data-Driven Differential Evolution with Multi-Level Penalty Functions and Surrogate Models” for a quintuplex extruder startup problem in which no explicit mathematical formulation of the objective or constraints is available. Historical production data are resampled to a common 1 Hz frequency, missing values are imputed using LOCF, and extrusion startups are segmented into optimization samples. Two surrogate models are then trained: a Random Forest classifier for minimum quality class and an XGBoost regressor for steadiness time. DE operates entirely on surrogate-based fitness, not on the real process (Garate-Perez et al., 15 Jul 2025).

The decision variables are derived from the stoppage period before extrusion initialization and include minimum, maximum, and mean values of pressure, minimum, maximum, and mean values of temperature, target screw speed, and mean startup acceleration of the screw. The optimization objective is steadiness time, defined procedurally from profilometer measurements using a 1% tolerance area and a 5-second sliding-window RMSE criterion. The first learned constraint predicts

$p=0.9$4

where class 1 is steady, class 2 not steady, and class 3 rejected before profilometer. This is incorporated through the multi-level penalty

$p=0.9$5

A second, deterministic physical consistency constraint uses

$p=0.9$6

with the consistency condition

$p=0.9$7

If violated, the final penalized objective becomes $p=0.9$8 (Garate-Perez et al., 15 Jul 2025).

The DE engine remains relatively standard, but initialization is also data-driven: 75% of the population is taken from best historical individuals and 25% from a fitted Gaussian Mixture Model over the empirical decision-space distribution. In this sense, search space, fitness, feasibility, and initialization are all informed by historical data. The reported optimized steadiness times are 9.93 s for product A, 4.10 s for product B, and 7.00 s for product C, compared with baseline predicted values of 24.82 s, 29.25 s, and 36.78 s; weighted by product distribution, the paper reports an average total reduction of 65% in the time needed to get quality steady (Garate-Perez et al., 15 Jul 2025).

5. Empirical behavior across benchmarks and applications

The empirical record of data-driven DE is favorable but not uniform. RevDE provides the clearest benchmark-dominance result among the population-geometry-driven methods. On Griewank, Rastrigin, Salomon, and Schwefel with $p=0.9$9, it is reported as best in all 12 benchmark cases in both convergence speed and final objective value. Its advantage is especially pronounced for $$\begin{aligned} \mathbf{y}_1 &= \mathbf{x}_i + F(\mathbf{x}_j - \mathbf{x}_k),\ \mathbf{y}_2 &= \mathbf{x}_j + F(\mathbf{x}_k - \mathbf{y}_1),\ \mathbf{y}_3 &= \mathbf{x}_k + F(\mathbf{y}_1 - \mathbf{y}_2). \end{aligned}$$0 and $$\begin{aligned} \mathbf{y}_1 &= \mathbf{x}_i + F(\mathbf{x}_j - \mathbf{x}_k),\ \mathbf{y}_2 &= \mathbf{x}_j + F(\mathbf{x}_k - \mathbf{y}_1),\ \mathbf{y}_3 &= \mathbf{x}_k + F(\mathbf{y}_1 - \mathbf{y}_2). \end{aligned}$$1. On the repressilator parameter-estimation problem, however, all methods reach almost identical final objective values and RevDE is only slightly faster on average. On the MNIST one-hidden-layer network with weight dimension 4120, ADE is best with test classification error $$\begin{aligned} \mathbf{y}_1 &= \mathbf{x}_i + F(\mathbf{x}_j - \mathbf{x}_k),\ \mathbf{y}_2 &= \mathbf{x}_j + F(\mathbf{x}_k - \mathbf{y}_1),\ \mathbf{y}_3 &= \mathbf{x}_k + F(\mathbf{y}_1 - \mathbf{y}_2). \end{aligned}$$2, RevDE is close at $$\begin{aligned} \mathbf{y}_1 &= \mathbf{x}_i + F(\mathbf{x}_j - \mathbf{x}_k),\ \mathbf{y}_2 &= \mathbf{x}_j + F(\mathbf{x}_k - \mathbf{y}_1),\ \mathbf{y}_3 &= \mathbf{x}_k + F(\mathbf{y}_1 - \mathbf{y}_2). \end{aligned}$$3, and both outperform DEx3 at $$\begin{aligned} \mathbf{y}_1 &= \mathbf{x}_i + F(\mathbf{x}_j - \mathbf{x}_k),\ \mathbf{y}_2 &= \mathbf{x}_j + F(\mathbf{x}_k - \mathbf{y}_1),\ \mathbf{y}_3 &= \mathbf{x}_k + F(\mathbf{y}_1 - \mathbf{y}_2). \end{aligned}$$4. RevDE is therefore not uniformly best across every task (Tomczak et al., 2020).

Clu-DE shows consistent improvement over plain DE on CEC-2017, but its evidence is relative to a basic baseline rather than to stronger adaptive DE families. The reported win/tie/loss counts are $$\begin{aligned} \mathbf{y}_1 &= \mathbf{x}_i + F(\mathbf{x}_j - \mathbf{x}_k),\ \mathbf{y}_2 &= \mathbf{x}_j + F(\mathbf{x}_k - \mathbf{y}_1),\ \mathbf{y}_3 &= \mathbf{x}_k + F(\mathbf{y}_1 - \mathbf{y}_2). \end{aligned}$$5 at $$\begin{aligned} \mathbf{y}_1 &= \mathbf{x}_i + F(\mathbf{x}_j - \mathbf{x}_k),\ \mathbf{y}_2 &= \mathbf{x}_j + F(\mathbf{x}_k - \mathbf{y}_1),\ \mathbf{y}_3 &= \mathbf{x}_k + F(\mathbf{y}_1 - \mathbf{y}_2). \end{aligned}$$6, $$\begin{aligned} \mathbf{y}_1 &= \mathbf{x}_i + F(\mathbf{x}_j - \mathbf{x}_k),\ \mathbf{y}_2 &= \mathbf{x}_j + F(\mathbf{x}_k - \mathbf{y}_1),\ \mathbf{y}_3 &= \mathbf{x}_k + F(\mathbf{y}_1 - \mathbf{y}_2). \end{aligned}$$7 at $$\begin{aligned} \mathbf{y}_1 &= \mathbf{x}_i + F(\mathbf{x}_j - \mathbf{x}_k),\ \mathbf{y}_2 &= \mathbf{x}_j + F(\mathbf{x}_k - \mathbf{y}_1),\ \mathbf{y}_3 &= \mathbf{x}_k + F(\mathbf{y}_1 - \mathbf{y}_2). \end{aligned}$$8, and $$\begin{aligned} \mathbf{y}_1 &= \mathbf{x}_i + F(\mathbf{x}_j - \mathbf{x}_k),\ \mathbf{y}_2 &= \mathbf{x}_j + F(\mathbf{x}_k - \mathbf{y}_1),\ \mathbf{y}_3 &= \mathbf{x}_k + F(\mathbf{y}_1 - \mathbf{y}_2). \end{aligned}$$9 at $\mathbf{Y}=\mathbf{R}\mathbf{X}$0, with repeated weakness on $\mathbf{Y}=\mathbf{R}\mathbf{X}$1. The paper also states that Clu-DE converges faster than standard DE on the plotted cases $\mathbf{Y}=\mathbf{R}\mathbf{X}$2 and $\mathbf{Y}=\mathbf{R}\mathbf{X}$3 (Mousavirad et al., 2021).

NbDE reports strong results on nine classical benchmark functions with dimensions 10, 30, and 50, often achieving the best rank and high success rates, but it does not dominate every case. DE/best/2 is better on $\mathbf{Y}=\mathbf{R}\mathbf{X}$4 and $\mathbf{Y}=\mathbf{R}\mathbf{X}$5 in some medium- and high-dimensional settings, and JADE is better on $\mathbf{Y}=\mathbf{R}\mathbf{X}$6 in several cases. The results support the claim that nearest-better guidance improves robustness against local optima, but not that it universally supersedes all alternatives (Dong et al., 2018).

MetaDE extends the empirical range beyond function optimization to Brax robotic control. It is reported to outperform or match representative DE variants on CEC2022 under both equal-time and selected equal-FE settings, and to outperform top CEC2022 entrants on many 20D functions. In Brax, it performs best on Swimmer but is slightly worse than CMA-ES and CSO on Hopper and Reacher. This reinforces a general pattern: data-driven DE methods tend to improve robustness and configurability, but their superiority is problem-dependent rather than absolute (Chen et al., 13 Feb 2025).

6. Limitations, misconceptions, and adjacent meanings

Several recurrent limitations follow directly from the cited studies. Population-data-driven methods do not necessarily save objective evaluations. RevDE enlarges the candidate set without extra sampling of source triplets, but the enlarged offspring still have to be evaluated eventually (Tomczak et al., 2020). Clu-DE adds one $\mathbf{Y}=\mathbf{R}\mathbf{X}$7-means clustering run per generation, yet provides no formal time-complexity analysis, no clustering sensitivity study, and no comparison against stronger baselines such as JADE, SHADE, or L-SHADE (Mousavirad et al., 2021). DE-DDQN requires a large offline training cost, uses handcrafted 99-dimensional state features, and learns only mutation-strategy choice while keeping $\mathbf{Y}=\mathbf{R}\mathbf{X}$8, $\mathbf{Y}=\mathbf{R}\mathbf{X}$9, $\det(\mathbf{R})=1$0, and restart logic fixed (Sharma et al., 2019). The extrusion study depends on surrogate quality; its XGBoost regressor has test Kendall correlation 0.41, provides no uncertainty quantification, and reports only offline validation rather than plant deployment (Garate-Perez et al., 15 Jul 2025). MetaDE relies on nested evaluations and a GPU-accelerated execution model; its one-shot fixed-seed evaluation reduces variance but does not establish robustness across stochastic realizations (Chen et al., 13 Feb 2025).

A common misconception is that any DE variant using population information is automatically “data-driven” in the strong machine-learning sense. Some works explicitly reject that interpretation. RevDE is data-driven relative to the population state, but it is not surrogate-assisted, not model-based DE, and not a learning-based operator-selection method (Tomczak et al., 2020). Similarly, DE with individuals redistribution is better described as feedback-driven or run-state-driven: it monitors low progress and diversity, then switches into a redistribution mode with standardized mutation, crossover, and selection-free trial retention (Li et al., 2020). At the opposite boundary, iStratDE is explicitly non-adaptive: mutation and crossover strategies are assigned once per individual at initialization and kept fixed for the full run. It is therefore a method based on static individual-level strategy diversity rather than direct data-driven control (Feng et al., 1 Feb 2026).

A second misconception is terminological. “Data-driven differential evolution” can be confused with “data-driven discovery of differential equations.” That is a different research area concerned with recovering governing ODEs and PDEs from data, often framed through sparse regression, symbolic regression, reinforcement learning, or neural equation discovery. A recent review on that topic explicitly states that it is not primarily about the Differential Evolution optimizer (Lou et al., 8 Jun 2026).

Taken together, the literature suggests that Data-Driven Differential Evolution is best treated as an umbrella concept with graded meanings. In its weakest form, it denotes DE variants that exploit population structure or run-time search statistics. In stronger forms, it denotes learned operator control, performance-driven meta-configuration, or surrogate-based optimization in which historical data define the effective objective and feasibility landscape. The core shift is always the same: DE ceases to be governed only by fixed stochastic rules and becomes conditioned, to varying degrees, on evidence extracted from data.