A stronger filter cannot invent the missing alternatives

Cross-Domain EchoesShared pattern · Natural Selection

A selection experiment can keep favoring a trait after the population’s response has largely stalled. One possible cause is loss of the inherited variation that previously supplied favorable alternatives. An evolutionary algorithm can encounter a related problem when strong selection narrows its candidate population faster than variation replenishes it. Both distinguish the pressure that favors alternatives from the process that makes alternatives available. The diagrams select this variation-limited case rather than explaining every biological plateau or stalled search. A continued filter can change which variants prevail without ensuring another step of improvement.

Written comparison

The available inherited alternatives

Quantitative genetics

Segregating genetic differences affecting the trait

Computational search

Encoded candidates and representation-compatible variation

Selection operates on alternatives present or generated in the relevant population.

The differential filter

Quantitative genetics

A persistent directional selection differential

Computational search

Evaluation-biased reproduction or survival

A nonzero preference does not itself establish that useful variation remains.

What crosses into the next round

Quantitative genetics

Inherited differences transmitted to descendants

Computational search

Candidate information transmitted through offspring

The filter can have cumulative effects because relevant information survives across rounds.

What can limit further response

Quantitative genetics

Depletion or fixation of favorable variation

Computational search

Premature loss of population diversity

These are selected failure modes, not the sole explanations of stalled progress or identical biological and computational laws.

What carries across

When progress stalls, inspect the available variation as well as the strength of selection. Repeated preference cannot select an alternative the process never supplies.

Where the comparison stops

A quantitative-genetic selection limit is not the same as an algorithm stopping or meeting a convergence tolerance. Distinct evidence is needed for each.

• Biological limits can also reflect fitness costs, linkage, drift, epistasis or physiological boundaries. The comparison selects depletion of useful variation.
• An algorithm’s variation operators can introduce or modify candidates, and stronger mutation is not automatically beneficial; representation, budget and evaluation still matter.
• No common rate of progress, guarantee of improvement, biological procedure or global-optimum conclusion follows.

Conditions for this comparison

• The biological plateau is sustained and distinguished from sampling noise or changed conditions.
• The computational case specifies selection pressure, representation and variation rather than treating all evolutionary algorithms as identical.
• Inherited retention is present in both cases; one-time sorting is insufficient.

Source entries

Shared pattern

Natural Selection

Prime

Core Idea

Natural selection is the structural engine in which *a population of differing variants is filtered by a pressure that lets the better-performing variants reproduce or persist more than the rest, so that — provided the differences are heritable — the population's composition shifts toward the favored variants over successive rounds*. Stated as a substrate-neutral schema, it has three irreducible ingredients and one consequence. First, *variation*: there is a population whose members differ from one another along some dimension that matters — phenotype, strategy, design, rule, belief. Second, *differential success under a selection pressure*: the variants do not all reproduce or persist equally; an environment or criterion confers on each variant a rate of reproduction or survival that depends on the variant's properties, so that some are favored and some are filtered out. Third, *heritable retention*: the properties that conferred success are *carried forward* into the next round — offspring resemble parents, surviving strategies are copied, retained designs seed the next generation — so the selection of one round biases the composition of the next. The consequence, when all three hold and the rounds repeat, is *cumulative adaptation*: the population becomes, over time, enriched in the variants the pressure favors, and can climb toward forms no single round's variation could have produced, because each round builds on the retained gains of the last.

Quantitative genetics

Selection limits

Domain-specific abstraction

Core Idea

In a breeding or selection experiment, the population mean initially shifts in the favored direction and later ceases sustained response even though the selection differential remains nonzero. Directional selection consumes favorable segregating alleles and can encounter fixation, mutation-selection balance, fitness costs, linkage, epistasis, drift, or physiological boundaries that reduce realized heritability.

Scope of Application

Selection limits belongs to quantitative genetics and is useful where the analyst can specify the typed quantitative genetics carrier, defining objects and relations, parameters, conventions, evidence, boundary cases, and comparison targets, then evaluate the trait, population, direction and strength of selection, generation history, response measure, effective size, and evidence distinguishing a true limit from noise or changed environment are explicit.

Computational search

Evolutionary Algorithm

Domain-specific abstraction

Core Idea

An evolutionary algorithm is a family of population-based stochastic search and optimization procedures inspired by variation, differential selection, and inheritance. It maintains computationally represented candidate solutions, evaluates their quality or behavior, uses selection to bias which candidates reproduce or survive, creates offspring with representation-compatible variation operators, and updates the population. The cycle continues until a budget, target, convergence test, or other termination condition is met.

Scope of Application

Evolutionary algorithms are used for black-box, discontinuous, noisy, mixed-variable, combinatorial, multimodal, and multiobjective problems where derivatives or exact solvers are unavailable or inconvenient. They also support design exploration, automated program construction, controller search, scheduling, feature selection, and model calibration. The family includes generational and steady-state population models, single- and multiobjective evaluation, fixed and self-adaptive operators, constrained optimization, coevolution, interactive evaluation, neuroevolution, and quality-diversity methods. Each extension must still bind representation, evaluation, selection, variation, and population update. An EA does not guarantee a global optimum merely because it explores stochastically. Performance depends on representation, operators, parameter settings, evaluation budget, and problem structure. A no-free-lunch boundary remains: success on a problem class reflects alignment between search bias and that class, not universal superiority.

Abstract Reasoning

Let \(P_t=\{x_1,\ldots,x_\mu\}\) be the population at iteration \(t\), and let \(E(x)\) be an evaluation. A generic cycle samples parents according to \(S_p(P_t,E)\), applies a variation kernel \(V(\cdot\mid\text{parents})\) to produce offspring \(O_t\), evaluates them, and uses survivor rule \(S_s(P_t,O_t,E)\) to form \(P_{t+1}\). This reveals two coupled distributions. Selection reweights the current population toward favored candidates; variation spreads mass into new candidates. Excessive selection pressure with weak variation can collapse diversity prematurely. Excessive variation or weak selection approaches random sampling. Search depends on balancing exploitation and exploration. Elitism—guaranteeing that selected high-quality candidates survive—can make the best-so-far objective nonworsening, but it does not imply the whole population improves or that the global optimum will be reached within a practical budget. Ergodic mutation plus infinite time can support asymptotic claims for some schemes, yet finite-run performance remains empirical and problem-dependent.