Skip to content

Differential Evolution

A population-based numerical optimizer that creates trial vectors from scaled differences among current candidates and retains them by evaluated quality.

Version
v1 · 2026-10-03 · History
Domain-specific #
13140
Domain group
Applied Sciences & Engineering
Origin domain
Computer Science & Software Engineering
Aliases
DE Optimizer, Differential Evolution Algorithm

Core Idea

Differential Evolution (DE) maintains a population of real-valued candidate vectors and forms new trials using scaled differences between current members. It evaluates each trial and selectively updates the population. Storn and Price's \(v=x_a+F(x_b-x_c)\) with binomial crossover is an original DE/rand/1/bin strategy, not the required equation or crossover rule of every variant.[ref-2f2fa5c58f5e][ref-599660d521bf]

Scope of Application

Storn and Price's constrained polynomial-coefficient benchmark was motivated by electronic filter design; it is not a report of building a physical filter. In a different setting, Stokes, Mandal and Wong used DE to optimize Arrhenius-model experimental temperature settings for a NO–ozone reaction. That output was a measurement design, not the direct fitted reaction-rate parameters.[ref-2f2fa5c58f5e][ref-ccdd87392d10]

Clarity

Identify the coordinates, bounds, objective, population, difference strategy, trial construction, selection and budget. DE's search does not require an objective gradient, but its evaluated criterion may use a model or derivatives internally. Do not assume a fixed crossover rate, automatic monotonically shrinking steps, universal noise robustness, or a global optimum from one run.[ref-2f2fa5c58f5e][ref-599660d521bf][^ref-ccdd87392d10]

Manages Complexity

Candidate differences supply proposal directions from the population's sampled geometry instead of an externally fixed random-step distribution. This can support broad search when members are spread apart and local proposals when they are close, but premature clustering can remove useful differences. Multiple objective calls and independent runs may still be costly.[^ref-2f2fa5c58f5e]

Abstract Reasoning

For current vectors \(x_i\), form a variant-specific trial from a weighted difference such as \(x_b-x_c\), evaluate its cost, and compare it with a target under the selection rule. The characteristic step is population-derived variation plus evaluated retention. A zero difference makes that particular displacement ineffective; a lower-cost replacement proves only a local comparison, not global optimality.[ref-2f2fa5c58f5e][ref-599660d521bf]

Knowledge Transfer

The polynomial and Arrhenius cases share vector population / differential proposal / trial / objective-guided update while changing what vector entries and cost mean. The staged strict parent is live domain-specific Evolutionary Algorithm, whose population/variation/selection cycle DE specializes. No canonical edge has been applied.[ref-2f2fa5c58f5e][ref-ccdd87392d10]

[^ref-2f2fa5c58f5e]: Rainer Storn and Kenneth Price, “Differential Evolution – A Simple and Efficient Heuristic for Global Optimization over Continuous Spaces,” Journal of Global Optimization 11 (1997), 341–359, §2 and §3.1 test 9. [^ref-599660d521bf]: SciPy developers, scipy.optimize.differential_evolution reference, v1.18.0, Notes and strategy formulas. [^ref-ccdd87392d10]: Zack Stokes, Abhyuday Mandal and Weng Kee Wong, “Using Differential Evolution to Design Optimal Experiments,” Chemometrics and Intelligent Laboratory Systems 199 (2020), 103955, §4.1.

Relationships to Other Abstractions

Local relationship map for Differential EvolutionParents appear above the current abstraction, mutual partners to the right, and children below. Node labels state whether each abstraction is prime or domain-specific; colors identify relation types.DifferentialEvolutionDOMAINDomain-specific abstraction: Evolutionary Algorithm — is a kind ofEvolutionaryAlgorithmDOMAIN

Current abstraction Differential Evolution Domain-specific

Parents (1) — more general patterns this builds on

  • Differential Evolution is a kind of Evolutionary Algorithm Domain-specific

    DE adds population-difference vector variation to an evaluated population/selection cycle.

Hierarchy paths (2) — routes to 2 parentless roots

Neighborhood in Abstraction Space

Differential Evolution sits in a sparse region of the domain-specific corpus (66th percentile for distinctiveness): few abstractions share its structure, so a faithful description tends to retrieve it precisely.

Family — Selection, Speciation & Experimental Evolution (22 abstractions)

Nearest neighbors

Computed from structural-signature embeddings · 2026-10-08