A new artificial-intelligence strategy is designed to tackle one of the most stubborn problems in computational science: how to find the best possible solutions when a system has thousands—or even millions—of variables, but only a small fraction of them actually matter. The method, developed by researchers at Shanghai Maritime University, uses an evolving “population” of candidate solutions that divides labor between two specialized groups. One group focuses on improving the quality of solutions already discovered, while the other searches for unexplored possibilities. By allowing these groups to evolve differently and by reallocating computing resources as the search progresses, the researchers report that their heterogeneous population co-evolutionary algorithm, or HPCEA, outperformed six leading algorithms across benchmark tests and three real-world scenarios.
The challenge arises in a class of problems known as sparse large-scale multi-objective optimization problems, or sparse LSMOPs. Optimization is the process of selecting the best configuration from many possibilities. Multi-objective optimization makes the task harder because “best” usually involves several competing goals at once: reducing cost while increasing efficiency, for example, or minimizing energy consumption while maintaining performance. In large-scale problems, the number of decision variables can become enormous. Sparsity adds another layer of difficulty: although the system may offer thousands of adjustable variables, the most effective solution may depend on only a small subset of them. The algorithm must therefore identify both which variables matter and what values they should take, while balancing multiple objectives that may conflict with one another.
Traditional evolutionary algorithms approach such problems by maintaining a population of candidate solutions and repeatedly applying operations inspired by biological evolution. Candidate solutions are evaluated, stronger individuals are selected, and new ones are generated through processes analogous to mutation and recombination. For multi-objective problems, an algorithm does not usually seek one single winner. Instead, it aims to approximate a Pareto front—the set of solutions for which improving one objective would worsen at least one other objective. A useful algorithm must converge toward this front while preserving diversity across it. If the population becomes too similar, the search may settle prematurely in one region. If it spreads too widely, it may fail to refine promising solutions. Sparse LSMOPs intensify both risks because irrelevant variables can introduce noise and consume much of the available computational effort.
According to the researchers, many existing co-evolutionary methods divide the problem into groups of variables or subpopulations using fixed rules. Such static arrangements can become inefficient as the search changes. Early in an optimization run, broad exploration may be essential because the algorithm has little information about the landscape. Later, concentrated refinement may be more valuable. A fixed allocation of individuals or computing time cannot easily respond to that transition. The result can be a mismatch between the algorithm’s resources and its evolutionary state: too much effort devoted to exploring when convergence is needed, or too much devoted to local improvement before the important regions of the search space have been found.
HPCEA addresses this mismatch with an adaptive population division strategy. Rather than permanently assigning a fixed number of candidate solutions to each task, the method adjusts the relative sizes of its convergence and diversity subpopulations according to real-time information from the evolutionary process. The convergence subpopulation is intended to push candidate solutions toward high-quality regions and improve their objective values. The diversity subpopulation has a different mission: it seeks alternative structures and helps prevent the search from collapsing into a narrow portion of the Pareto front. The precise balance between these populations can therefore change as the search evolves, allowing computational resources to be directed toward whichever function is most urgently needed.
The algorithm’s second major component is a heterogeneous approach to grouping variables. In the convergence subpopulation, HPCEA uses correlation grouping to identify variables whose behavior appears structurally linked. Correlation, in this setting, can reveal relationships among decision variables or between variables and the optimization landscape. Variables that influence one another may be handled together so that the algorithm can make coordinated changes rather than treating every dimension independently. This matters in high-dimensional spaces because independent manipulation of related variables can destroy useful combinations. By recognizing patterns of association, the convergence-oriented search can focus its effort on coordinated refinements instead of blindly perturbing an overwhelming number of dimensions.
The diversity subpopulation uses a different strategy called structural grouping. Its purpose is not simply to exploit correlations already visible in the current population, but to uncover potential sparse patterns that may not yet be obvious. In a sparse problem, the algorithm must decide which variables should be active and which should remain effectively inactive. Structural grouping provides a way to investigate possible arrangements of active variables, helping the search discover alternative solution architectures. This division of labor is central to HPCEA’s design: one population uses information about known relationships to accelerate convergence, while the other probes for new patterns that could expand the range of viable solutions.
HPCEA also assigns sparse importance scores to variables and uses heterogeneous genetic operators to manipulate them. In evolutionary computation, genetic operators are the mechanisms that create new candidate solutions from existing ones. Mutation may alter selected variables, while crossover or recombination may combine information from multiple candidates. Applying the same operator uniformly across every variable can be wasteful in a sparse, high-dimensional problem. Variables believed to be important may need careful, targeted adjustment, whereas low-importance variables may be suppressed, activated experimentally, or altered more aggressively to test whether they contain hidden value. By combining importance estimates with the different structures of its two subpopulations, HPCEA is designed to balance exploitation—rapid improvement of promising solutions—with exploration of uncertain regions.
The researchers evaluated the method against six state-of-the-art algorithms using eight benchmark suites and three real-world scenarios. The study reports that HPCEA showed superior overall performance, although the supplied research summary does not provide the individual benchmark scores, the names of the competing algorithms, or details of the real-world applications. The result is nevertheless significant because sparse LSMOPs are increasingly relevant to systems in which models contain many possible controls but only a few should be active. Examples can include engineering design, resource allocation, network configuration and other decision problems where efficiency depends on selecting a compact set of influential variables. An algorithm that can identify those variables while preserving a broad range of trade-off solutions could reduce the computational burden of designing and managing such systems.
The work does not eliminate the fundamental difficulty of large-scale optimization; instead, it changes how the search is organized. Its key idea is that a single homogeneous population should not be expected to perform every task equally well. Convergence and diversity require different kinds of information, different variable groupings and potentially different genetic operations. HPCEA turns that observation into an adaptive architecture in which populations specialize, exchange evolutionary pressure and receive resources according to the state of the search. The authors say the study demonstrates the method’s superiority across extensive experiments, but further work would be needed to determine how it behaves under noisy objectives, changing constraints, extremely high-dimensional variables or computational limits outside the tested scenarios. For now, the algorithm offers a striking example of how borrowing the language of evolution can lead to a more specialized form of machine intelligence—one capable not only of searching for better answers, but also of deciding which parts of a problem deserve attention.
Subject of Research: A heterogeneous population co-evolutionary algorithm for sparse large-scale multi-objective optimization problems
Subject of Research: Technology and Engineering
Article Title: A heterogeneous population co-evolutionary algorithm for sparse large-scale multi-objective optimization problems
Article References: Wang, C., Liu, T., & Yang, G. (2026). A heterogeneous population co-evolutionary algorithm for sparse large-scale multi-objective optimization problems. Complex & Intelligent Systems. https://doi.org/10.1007/s40747-026-02409-x
Image Credits: AI Generated
DOI: 10.1007/s40747-026-02409-x
Keywords: sparse optimization, large-scale multi-objective optimization, co-evolutionary algorithm, heterogeneous population, Pareto optimization, evolutionary computation, variable grouping, computational intelligence
Cite this news
APA MLA Chicago
SCIENMAG. (August 28, 2026). New Co-Evolutionary Algorithm Tackles Sparse, Large-Scale Multi-Objective Optimization Problems. https://scienmag.com/new-co-evolutionary-algorithm-tackles-sparse-large-scale-multi-objective-optimization-problems/
SCIENMAG. “New Co-Evolutionary Algorithm Tackles Sparse, Large-Scale Multi-Objective Optimization Problems.” Scienmag, 28 August 2026, https://scienmag.com/new-co-evolutionary-algorithm-tackles-sparse-large-scale-multi-objective-optimization-problems/. Accessed 28 August 2026.
SCIENMAG. “New Co-Evolutionary Algorithm Tackles Sparse, Large-Scale Multi-Objective Optimization Problems.” Scienmag. August 28, 2026. https://scienmag.com/new-co-evolutionary-algorithm-tackles-sparse-large-scale-multi-objective-optimization-problems/
Copy citation Download RIS
Tags: artificial intelligence in optimizationbenchmark testing of evolutionary algorithmsbenchmark testing of optimization algorithmsco-evolutionary algorithmsevolutionary algorithms for sparse dataheterogeneous population in AIheterogeneous population optimizationHPCEA for multi-objective optimizationlarge-scale multi-variable optimizationlarge-scale variable optimizationmulti-objective evolutionary algorithmsmulti-objective problem solvingmulti-objective problem-solving in AIoptimizing multi-objective systems with sparse variablesreal-world applications of HPCEAreal-world multi-objective applicationsresource allocation in optimizationscalable multi-objective optimization techniquesscalable optimization algorithmssparse large-scale multi-objective optimizationsparse system solution strategiesspecialized co-evolutionary strategies
