Multi-Population Evolutionary Algorithms

• Multi-Population Evolutionary Algorithms are a class of methods that evolve several distinct populations in parallel, each potentially employing unique strategies to explore different search spaces.
• They utilize structured migration, selection, and replacement mechanisms to balance rapid convergence with the preservation of diversity, ensuring robust global search performance.
• These algorithms are widely applied in global optimization, neural architecture search, multi-objective optimization, and adversarial training, often outperforming single-population approaches.

A multi-population evolutionary algorithm (MPEA) is a class of evolutionary algorithms (EAs) in which more than one population (colloquially, “swarm,” “island,” “archive,” or “deme”) evolves in parallel, with structured mechanisms for information exchange. The fundamental aim is to enhance exploration, maintain diversity, avoid premature convergence, and, in many use cases, efficiently leverage parallel computing resources. Each population may have different roles, employ distinct operators, occupy different regions of the search or objective space, or be connected via explicit migration policies. Multi-population methods are prominent in real-valued global optimization, neural architecture search, multi-objective optimization, and evolutionary learning in adversarial settings.

1. Formal Definitions and Canonical Architectures

Let $M$ denote the number of populations. For each population $P_m$ ($m=1,...,M$), the set of individuals $\{x_{i}^{(m)}\}$ is evolved using evolutionary operators (selection, crossover, mutation), and a fitness function $F^{(m)}$. MPEA architectures can be classified as follows:

• Synchronous island model: Populations (islands) evolve independently with occasional migration events, typically exchanging elite or random individuals with spatial/temporal synchronization. The connectivity is characterized by an adjacency matrix $A$.
• Asymmetric or bi-population frameworks: Populations have differentiated roles—e.g., exploitation versus exploration, or focused on feasibility versus diversity (Li et al., 2017, Xue et al., 9 Feb 2026).
• Layered/partitioned mechanisms: Distinct subspaces of the decision or search space are assigned to different populations, sometimes in a decomposition-based framework (Huang et al., 24 Sep 2025).
• Co-evolutionary setups: Multiple populations represent antagonistic or co-adapting entities, such as generators and discriminators for GAN training (Casas et al., 17 Jul 2025).

Each population is updated according to its evolutionary dynamics, and information exchange (migration, recombination, or elitism) is governed by problem-dependent schemes. For constraint-handling, objective-decomposition, or parallel hardware, the populations may use specialized selection or fitness assignment.

2. Selection, Replacement, and Interaction Mechanisms

MPEAs employ diverse mechanisms for selection, survivor replacement, and inter-population interaction:

• Selection and Replacement: Classical schemes include ($\mu,\lambda$)-full generational replacement, ($\mu+\lambda$)-elitism, or tournament selection (Casas et al., 17 Jul 2025). In bi-population and archive-based algorithms, restricted mating selection dynamically allocates parent selection probabilities between populations, conditioned on the current dominance structure (Li et al., 2017, Xue et al., 9 Feb 2026).
• Migration: Defined by migration intervals ($\Delta t_{mig}$), migration fractions ($r_{mig}$), and topology ($P_m$0). Policy specifics can involve migrating elite individuals, random individuals, or hybrids, with symmetric or asymmetric flows (Federici et al., 2020, Schiavello et al., 2024).
• Diversity-driven mechanisms: Epidemic restarts trigger partial population reinitialization when diversity drops below a threshold, preserving only elite subsets (Federici et al., 2020).
• Role-differentiated populations: For example, constraint-oriented and unconstrained subpopulations in constrained multi-objective optimization (CMOP) (Huang et al., 24 Sep 2025), or convergence and diversity archives in two-archive EAs (Li et al., 2017).

The following table summarizes common interaction schemes:

Architecture	Migration Policy	Main Purpose
Island Model	Periodic, topology-$P_m$1	Parallel search, diversity
Bi-population	Elite, asymmetric	Coverage of search regions
Co-evolutionary	Generational sync	Red/blue team dynamics
Archive-based	Restricted mating	Feasibility versus diversity

3. Algorithmic Instantiations and Pseudocode

Synchronous Island Model (EOS (Federici et al., 2020)):

$P_m$2

Migration alternates between inward and outward flows, with best individuals sent from each island to neighbors, replacing the local worst. Each island may utilize distinct DE mutation strategies (e.g., explorative versus exploitative).

Dual Population for Neural Architecture Search (MOEA-BUS (Xue et al., 9 Feb 2026)):

$P_m$3

Co-evolutionary GANs (full generational replacement (Casas et al., 17 Jul 2025)):

$P_m$4

4. Theoretical Analysis: Diversity, Convergence, and Exploration

Statistical mechanics approaches provide analytical models for the evolution of population-level statistics under migration, selection, and stochastic operators (Messias et al., 2018). After migration, the cumulants of each island population's fitness mix as weighted averages, quantifying diversity retention and homogenization rates:

$P_m$5

Selection strength, migration rate, and topology directly impact the trade-off between rapid convergence and diversity maintenance. Strong selection ($P_m$6) or excessive migration can precipitate premature convergence, while sparse connectivity and lower migration rates preserve exploratory power but may decelerate global search (Messias et al., 2018).

In co-evolutionary settings (e.g., GANs), full generational replacement strategies (($P_m$7)) prevent population monopolization, maximize exploration, and outperform elitist or tournament-based replacements in sample quality and diversity metrics (Casas et al., 17 Jul 2025). In bi-/multi-population NAS, asymmetric migration and uniform initialization amplify entropy and hypervolume on the Pareto front (Xue et al., 9 Feb 2026).

5. Applications in Constrained, Multi-objective, and Adversarial Settings

• Constrained Multi-objective Optimization: GMPEA (Huang et al., 24 Sep 2025) uses decomposition-based dual-populations, one focusing on constraint satisfaction and the other promoting unconstrained global search. This leads to competitive IGD and hypervolume metrics under strict time constraints, and achieves up to $P_m$8 GPU speedups relative to state-of-the-art CPU and partially tensorized GPU baselines.
• Neural Architecture Search: Bi-population (MOEA-BUS (Xue et al., 9 Feb 2026)) and archive-based (SMEM-NAS (Xue et al., 2024)) approaches promote search-space coverage and robust Pareto diversity, outperforming single-population NAS in terms of accuracy and complexity trade-offs.
• Quantum-classical Hybrid Optimization: Distributed evolutionary QAOA (Schiavello et al., 2024) employs parallel populations on multiple quantum processing units, exchanging elite individuals. This method consistently exceeds single-population/gradient-based baselines in Max-Cut approximation ratios while maintaining higher diversity and convergence speed.
• Co-evolutionary Adversarial Training: Dual-population frameworks stabilize adversarial learning by evolving separate populations of generators and discriminators, with empirical superiority