Dynamic Population Sizing for Efficient Evolutionary Search

This direction is merged into another direction: dynamic population sizing. Its last version stays readable and citable here.

Version 1 of 123 papers31 conceptsfrozen 2026-10-04

In evolutionary computation, the size of the population is often the most critical hyperparameter. Set it too small, and the algorithm suffers from "premature convergence," where the population loses genetic diversity and gets stuck in a local optimum before finding the global solution. Set it too large, and you waste massive amounts of computational power on redundant fitness evaluations that offer diminishing returns.

Traditionally, researchers have treated population size as a static value to be tuned through trial and error. However, a growing body of research suggests that the "optimal" population size is not a fixed number, but a moving target that should change as the search progresses.

The Exploration Exploitation Tradeoff

The fundamental intuition behind dynamic population sizing is the shifting requirement of the search process. At the beginning of a run, a large population is beneficial for exploration—covering as much of the search space as possible to identify promising regions. As the algorithm converges on a solution, the focus shifts to exploitation, where fewer individuals are needed to refine the result.

Juan Luís Jiménez Laredo and Christian Gagné formalized this in their work on Simple Variable Population Sizing (SVPS). Their approach uses a deterministic shrinkage scheme: start with a large population to ensure diversity and systematically reduce the number of individuals over time. This allows the algorithm to maintain the success rate of a large-population GA while significantly cutting the total number of fitness evaluations.

Growing to Solve Complexity

While shrinkage helps with efficiency, some problems require the opposite: starting small and growing. In multi-objective optimization, the goal is to find a "Pareto front"—a set of optimal trade-offs between conflicting goals.

Simon Wietheger, Benjamin Doerr, and Martin S. Krejca demonstrated in Speeding Up the NSGA-II via Dynamic Population Sizes that starting with a tiny population and gradually increasing it can break the theoretical "speed limit" of multi-objective solvers. By scaling the population size in proportion to the discovered complexity of the Pareto front, they achieved a speedup proportional to the problem size itself.

Performance-Driven Feedback Loops

More advanced methods move away from deterministic schedules (like simple shrinkage or growth) and instead use the state of the search to decide when to scale. This creates a feedback loop where the algorithm "breathes" based on its progress:

  • Stagnation-Triggered Growth: Iana S. Polonskaia and Nikolay Nikitin proposed using the Fibonacci sequence to adjust populations in the FEDOT framework. If the search stagnates, the population grows to inject new diversity; if improvements are found, it shrinks to conserve resources.
  • Niche Balancing: In multimodal optimization, where the goal is to find multiple different solutions, Ting Huang and Sam Kwong developed a Niching Memetic Algorithm that adjusts the size of sub-populations. High-performing "niches" are pruned to save costs, while underperforming ones are expanded to encourage deeper exploration.
  • Biological Metaphors: Junwan Liu and Xiaohua Hu applied dynamic sizing to artificial immune systems and "shuffled frog-leaping" algorithms (MODPSFLB). Their systems expand the population to investigate new data regions and contract it by "pruning" redundant or low-performing agents.

Eliminating the Human in the Loop

The ultimate goal of this research direction is to remove the need for manual parameter tuning entirely. Anton Bouter and Peter Bosman have integrated these concepts into the gomea library using interleaved multi-start schemes.

Instead of picking one size, the algorithm manages several concurrent populations of different sizes simultaneously. By dynamically allocating more resources to the populations that show the most promise, the system becomes robust to different problem difficulties without any user intervention. As Youhei Akimoto noted in the development of AREX, providing this autonomy is essential for solving complex, real-world landscapes like ridges and multi-modal functions where a single fixed population size is almost never optimal.

Papers in this version

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_dynamic_population_sizing_for_efficient__5eaa1831,
  title = {Dynamic Population Sizing for Efficient Evolutionary Search},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/dynamic-population-sizing-for-efficient-evolutionary-search/rdr_5eaa1831b8bf415f9fc16dc5e6144a0a},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_dynamic_population_sizing_for_efficient__5eaa1831_v1,
  title = {Dynamic Population Sizing for Efficient Evolutionary Search},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/dynamic-population-sizing-for-efficient-evolutionary-search/rdr_5eaa1831b8bf415f9fc16dc5e6144a0a/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}