Early work examined diploid representations and diversity in changing environments [12], [14], [15]. Later memory and multi-population methods stored useful solutions or maintained search in multiple regions [13], [16], [21]. These sources establish antecedents for AMPERE’s modules; they do not establish that the particular combined implementation is superior.

Dynamic test generators and moving-peaks configurations provide controlled ways to study changing fitness landscapes [13], [19], [20]. Surveys distinguish memory, diversity and prediction mechanisms [17], [22]. The reported campaign uses measures of current-best fitness and recovery together with a within-method incumbent-drop measure. The latter differs from benchmark error computed relative to a known current optimum, so it should not be read as an interchangeable measure of tracking accuracy.

Operator adaptation and uncertainty handling are also covered in the dynamic-optimization literature [13], [18]. Their relevance depends on the change process and evaluation constraints. The present transition estimate represents only the reported small regime space and does not demonstrate accurate prediction across broader environment classes.

Chemostat growth and inhibition models provide one basis for interpreting nutrient-dependent dynamics [1– 3], [23]. Classical population models and experiments [24], [25] and later microbial-interaction studies [11], [26–28] concern ecological structure and stability. The present experiment uses a fixed two-species surrogate with continuous candidate variables; it does not optimize community membership or certify ecological stability. Genome-scale and dynamic flux-balance methods [29], [30] remain related modeling approaches rather than executed components.

Phenotypic switching, bet-hedging and persistence provide biological reasons to study fluctuating environments [6], [7], [31]. These mechanisms concern organisms and populations. They motivate the application setting, but do not validate the numerical response of an evolutionary search algorithm or establish that simulated design changes correspond to realizable strain adaptation. Here the species set is fixed and a stylized feed-matching trait is an optimizer variable.

Bioprocess estimation and control address changing states and uncertain kinetics [8], [9]. Robust optimization also covers multiple uncertainty settings and cannot be restricted categorically to a static environment [32]. These approaches are plausible comparators, but the present campaign includes neither a robust-design optimizer nor model-predictive control. Metabolic-community and dynamic flux-balance models offer richer biological representations [29], [30]. However, the reported chemostat surrogate does not execute a genome-scale flux solve. ANN–metaheuristic optimization has also been applied to enzyme and biomass responses as functions of fermentation-medium composition [44]. That static single-organism response problem differs from the changing two-species dynamical surrogate considered here.

The relevant research question is how search mechanisms behave when fitness is generated by a changing simulator. Table 1 summarizes potential roles of three research traditions rather than asserting that any tradition lacks an entire capability. The numerical comparison is restricted to the configured methods and models. It neither demonstrates an exhaustive literature gap nor measures an industrial advantage.

contrast and model class independently.

Modelling Model-dependent Optimization can design studies support control Interactions and Model-dependent growth mechanisms process dynamics Variable-environment Estimation and analysis disturbance handling Ecological and Predictive control evolutionary models where specified Model and state Estimated states and histories controller memory

The executed nutrient process is described as a discrete two-state Markov chain [33] at scheduled optimizer epoch boundaries, with a separate fine-step feed chain inside each 96-hour simulation. It is not a documented continuous-time jump-process implementation. The species set is fixed at two; the exact continuous design coordinates, bounds, initial states and kinetic constants must be recovered from the original simulator. Table 2 separates physical-time quantities from optimizer-indexed quantities.

z_e macro nutrient regime at optimizer epoch e dimensionless Z finite regime set dimensionless P discrete macro-regime transition matrix dimensionless x_i(t) biomass concentration of species i g L⁻¹ s_k(t) substrate concentration k g L⁻¹ μ_i(s; z) specific growth rate h⁻¹ θ continuous candidate design; species set fixed coordinate-dependent f(θ; z) 96-hour endpoint titer in the chemostat g L⁻¹ E_off within-method incumbent drop averaged over optimizer steps fitness units τ_rec censored first crossing of the pooled epoch reference optimizer generations

Eq. (1) gives a generic biomass balance for interpreting a chemostat model, rather than a complete specification of the executed simulator. A full model also requires substrate, by-product and product balances, uptake functions, yields, initial states and the decision-to-parameter map. Eq. (2) defines the reported fitness as endpoint product concentration. Eq. (3) states the discrete macro-regime transition model; a symmetric switching probability of 0.6 corresponds to off-diagonal entries of 0.6 and diagonal entries of 0.4. Here x_i is biomass concentration, D is dilution rate and k_d,i is a loss or decay rate, with both rates in h⁻¹. P_{θ,z}(t) denotes product concentration in g/L for decision θ and regime z; this scalar concentration is distinct from the transition matrix P in Eq. (3).

dxi

= [μ (s;z) − D − k ]x . (1)

i d,i i

dt

f(θ; z) = Pθ,z(96 h). (2)

Pr(ze+1 = j ∣ ze = i) = Pij, Pij ≥ 0, ∑ Pij = 1.

For a theoretical tracking objective, let π be a causal search policy producing a design θg at optimizer index g, with access only to observations available by that index. Let f*(z) be the supremum of the expected endpoint fitness over the stated design box in regime z. Eq. (4) gives an idealized expected gap for this policy. This comparator is not computed by the reported campaign, whose finite probes do not certify f*(z).

min E[ ∑(f (zg) − f(θg ; zg))]. (4)

Eq. (4) optimizes a policy rather than a single fixed design. It contains a fitness-gap term and no switching-cost term. Simulation time inside a fitness evaluation and optimizer index across evaluations serve different roles. The reported mean current-best titer averages optimizer outcomes from reset simulation windows; it is not time-averaged production from one continuously operated reactor.

The reported detector re-evaluates five fixed sentinel designs each generation and combines a two-sided 5 percent fitness-displacement rule with cumulative-sum monitoring [34]. The supplied description does not resolve the CUSUM threshold, warm-up, reset policy or whether the two rules are joined by conjunction, disjunction or a confidence-dependent gate. Detection results are therefore reported without attributing missed events to a particular rule.

The observations in this campaign are simulator fitness values. Online substrate, off-gas or product measurements from a biological process were not tested. The deterministic evaluation setting does not establish robustness to measurement noise or quantify the separate benefit of cumulative-sum monitoring.

The reported memory contains regime-indexed elites and a recent-elite fallback. Stored designs are re-evaluated before reuse. The transition estimate uses observed regime transitions with a Dirichlet prior. Exact elite capacity, regime-assignment rule, count update and reset conventions are implementation dependencies; nearest-centroid classification and fitted dwell-time distributions are not documented executed components. Forgetting may be useful when stored solutions become stale, but exponential discounting of transition counts does not by itself bound the number of archived regimes or designs. A finite capacity and deletion policy would be required for a memory bound. Such a policy should be specified from the source code before any bounded-memory claim is made.

The reported allocation assigns search effort using estimated successor probabilities and activates at a successor probability of 0.25. Population sizes, floors, migration and the handling of counterfactual-regime evaluations need an explicit implementation record. Cached regime parameters do not make new fitness evaluations free; their simulation calls must be charged to the same budget.

Allocation can be viewed as a choice between current tracking and preparation for possible future regimes. Its benefit depends on prediction quality and the cost of diverted evaluations. No-free-lunch results [35] do not prove a benefit or a failure on either restricted benchmark family; those questions require the relevant controlled comparisons.

The reported mutation response increases with detected change magnitude and is capped when re-evaluated elites remain competitive. Its exact scaling function, restart fraction and cap require source-code verification. The fixed-hypermutation comparator multiplies the mutation scale by ten for ten generations, providing a distinct configured response rather than an otherwise identical controller.

The executed GA uses tournament selection of size three, simulated binary crossover with index 15 and polynomial mutation with index 20, without elitism. There is no reported strain-membership bit string in this fixed two-species experiment. The operational sequence below is a summary of the reported modules, not a replacement executable specification.

The search begins with the candidate population, sentinel designs and empty or initialized memory records. Within an epoch, it evaluates and varies designs under the active fitness configuration while updating sentinel statistics. Upon a detected change it updates available regime records, re-evaluates reusable elites, applies the configured mutation response and updates allocation. Execution stops at the shared evaluation budget; implementation-dependent ordering, accounting and resets require the original code.

For population size N and s sentinel designs, with a denoting the number of archive re-evaluations at each change, the immediate evaluation overhead is s calls per monitored generation and a calls per archive refresh. Total runtime also depends on simulator cost and counterfactual evaluations. No logarithmic archive retrieval or bounded-growth guarantee follows from the supplied description. Matched total evaluation limits can constrain calls without making offspring counts or generation counts identical.

Memory and anticipation require reusable regime structure and informative observations. Noise, misclassification and changing kinetics can degrade both modules. The reported campaign does not independently manipulate these factors, so it cannot identify a recurrence threshold or a noise level at which AMPERE becomes preferable.

False-negative and false-positive detection rates require an event-level record with a defined denominator, tolerance window and treatment of repeated alarms. Reported detection totals alone do not provide both rates.

Unbounded or continuously changing regime spaces also require a specified clustering and archive-capacity policy. Simulator–process mismatch remains a separate external-validity issue and was not measured in either benchmark.

The reported comparison uses six population-based methods. Moving peaks has five dimensions on [0,100]⁵, ten sphere peaks with heights initially uniform on [30,70] and widths on [1], [5], shift severity 1.0 and a clipped Gaussian height walk of severity 7.0. Changes occur every 2,500 evaluations; each run has 25 epochs and 62,500 evaluations, with 24 scheduled changes and population size 50. The reported configuration is non-recurrent; its exact generator and peak-update conventions remain necessary for reproduction [13], [19]. The chemostat surrogate fixes a producer and a consumer. The producer forms an inhibitory by-product and product; the consumer removes by-product while incurring maintenance and product-consumption costs. A stylized producer trait has different preferred values under the two macro regimes. The reported macro-state chain switches with probability 0.6 at scheduled boundaries. Inside each reset 96-hour simulation, feed switching is approximately 0.001 per 0.5-hour step. This corresponds to about 0.002 switches per hour, or a discrete-step hazard −log(0.999)/0.5 ≈ 0.002001 h⁻¹, rather than a continuous-time intensity of 0.001 h⁻¹. Explicit Euler uses 192 steps with positivity clipping. Fitness is endpoint titer in g/L. Seven optimizer epochs give 17,500 evaluations and six scheduled boundaries per seed; a boundary need not produce an actual macro-state switch.

The reported baselines are a generational GA with tournament size three, crossover index 15, mutation index 20 and no elitism; random immigrants replacing 20 percent of offspring [14], [15], memory-enhanced search re-evaluating ten recent epoch elites and reinserting the five best [13], [36], triggered hypermutation multiplying mutation scale tenfold for ten generations [15], and five subpopulations of ten with ring exchange every five generations [16]. AMPERE adds the reported regime archive, Dirichlet-regularized transition estimate, allocation threshold of 0.25 and magnitude-scaled response. Detection-equipped methods share the stated five-sentinel detector, but other module and accounting differences remain relevant to attribution. The source reports five sentinel calls per monitored generation, ten archive re-evaluations per change, common environmental realizations within each main-comparison seed, ten main seeds, and five ablation seeds. Exact counters are needed to verify that all calls, including successor-regime tests, fall within the budget. NumPy [37], SciPy [38] and Matplotlib [39] are reported, with a total runtime of 304 seconds on one core; the original run log is needed to attest that measurement. A 400-point Latin-hypercube probe produced reported best titers of 4.09, 4.08, 4.07, and 3.99 g/L, as well as trait values of 1.29 and 0.79 in the contrasted configurations. These are finite-sample observations, not proven regime optima or ceilings.

For method m and epoch e, let bmg be its current-best fitness at optimizer step g and hmg = max{bmu: u is in the same epoch and u ≤ g}. The reported offline-error column averages hmg − bmg. It therefore measures a loss relative to that method’s own running incumbent, not to a known optimum, and can remain zero for consistently poor designs. The final-window error uses the last five generations. Recovery is the first generation reaching 95 percent of the best value attained by any method in the same seed–epoch pool; non-attainment is recorded at the epoch length. Such a censored, retrospectively pooled measure is not an uncensored expected hitting time. Because overhead changes the number of generations, evaluation-index recovery would be needed for a stricter budget comparison.

Table 3 reports the moving-peaks summaries, and Fig. 1 shows a representative aggregate trajectory. The plain GA has the smallest incumbent-drop metric (0.103 ± 0.031), followed by memory-enhanced and multi-population GAs (0.125 ± 0.056 and 0.156 ± 0.055). AMPERE gives 0.293 ± 0.124, below random immigrants (0.557 ± 0.415) and fixed hypermutation (2.622 ± 0.788). These differences are compatible with disruptive variation under small shifts, but aggregate traces do not isolate that mechanism. Multi-population search has the largest reported mean current-best fitness, 53.1 ± 8.7, and the fastest pooled-reference recovery, 24.1 ± 8.0 generations. Thus the incumbent-drop ordering is not the complete objective-quality ordering. The detector reports 219 detections among 240 scheduled changes, corresponding to 91.25%; event logs are required to establish missed-event causes or false-alarm rates.

Table 4 reports component ablations. On moving peaks, disabling the mutation response gives an incumbent-drop metric of 0.111 ± 0.075, compared with 0.293 ± 0.124 for the displayed full method. The reported paired W = 1.0 and p = 0.125 do not establish improvement at the 0.05 level or agreement in every nonzero pair. Removing the archive gives 0.424 ± 0.244. On the chemostat, disabling the response gives 1.76 ± 0.50 g/L rather than the displayed 4.08 ± 0.17 g/L, with W = 0 and p = 0.0625. For five nonzero paired differences under the exact sign-randomization null hypothesis, the smallest two-sided p-value is 2/2⁵ = 0.0625. Archive, allocation and transition ablations have displayed means 4.14 ± 0.02, 4.15 ± 0.02 and 4.15 ± 0.01 g/L, without a significant reported contrast. This does not establish equivalence, redundancy, or a uniquely causal module effect. The full rows duplicate main-comparison summaries; the matching five-seed full-method outputs must be supplied to verify the paired ablation statistics.

Table 5 reports the chemostat summaries. AMPERE’s mean current-best endpoint titer is 4.08 ± 0.17 g/L, while random immigrants and triggered hypermutation give 4.14 ± 0.01 and 4.12 ± 0.02 g/L. Plain, multi-population and memory-enhanced GAs give 1.55 ± 0.41, 1.55 ± 0.42 and 1.62 ± 0.41 g/L. The finite-probe value of 4.10 g/L is not a ceiling: two displayed method means exceed it. The source reports recovery censoring at 36, 36 and 35 of 60 events for the latter methods, and detection totals of 40/40 for detector-equipped methods. These denominators cannot be identified with one another without an event log distinguishing scheduled boundaries, actual macro switches and eligible detection or recovery events. No failure mechanism is inferred from these totals alone.

The two benchmarks give different configured rankings, but differ in budget, geometry, fitness scale and regime construction as well as recurrence. The comparison therefore does not isolate recurrence or establish quantitative transfer. On the chemostat, low within-method drop values coexist with low titers, so inference uses titer rather than that drop metric. The reported comparisons do not isolate an additional benefit of transition prediction or regime-indexed memory beyond diversification.

Table 6 preserves the reported Friedman and paired signed-rank summaries [40–43]. The moving-peaks Friedman statistic 41.2 with five degrees of freedom has an asymptotic chi-square tail of approximately 8.55 × 10⁻⁸; the displayed 0.0000 is rounded, not zero. AMPERE has larger incumbent-drop values than the plain and memory-enhanced GAs (Holm p = 0.0098 and 0.0176), but smaller values than triggered hypermutation (Holm p = 0.0098). Its contrast with multi-population search does not reject at 0.05 (Holm p = 0.0547), and it is numerically better than random immigrants with an unadjusted p = 0.1934. For the chemostat, the source reports chi-square = 40.7 and p = 1.09 × 10⁻⁷; using the rounded statistic gives approximately 1.08 × 10⁻⁷. AMPERE has higher titer than the plain, memory-enhanced and multi-population GAs (Holm p = 0.0098 for each). The random-immigrant and triggered-hypermutation comparisons do not reject; this is not evidence of equivalence. Main-test pairing, tie handling, correction families and the five-seed ablation reference require the original arrays. Seeds from a single parameterized instance do not represent ten independently sampled biological systems.

The reported pattern depends on the benchmark. AMPERE loses to the plain and memory-enhanced GAs on the moving-peaks incumbent-drop metric, while the multi-population GA leads mean current-best fitness. On the chemostat, AMPERE and two simpler diversification methods have higher reported mean endpoint titers than three less disruptive search methods. The mutation-response ablation is consistent with a useful response in this constructed chemostat, but its five-pair result is not significant at 0.05 and does not prove a unique mechanism. Memory, transition modeling and allocation remain unisolated rather than proven redundant.

The detector and archive consume fitness calls that could otherwise produce offspring. Their fraction depends on the realized generation and change counts; a universal 10 percent overhead is not established. The two fitness maps are reported as deterministic, so robustness to noisy biological observations remains untested. The two-species surrogate uses a deliberately contrasting trait and a coarse clipped Euler integrator. Integration convergence, parameter sensitivity and validation against measurements are needed before biological or process-level conclusions can be made.

These observations motivate a controlled follow-up that varies regime recurrence and contrast independently, uses matched evaluation accounting, and includes separate ablations for detection, memory, prediction, and mutation. They do not identify a general operating rule for an industrial bioprocess. A practically useful controller would additionally require representation of feasible actions, sensing delays and costs, state continuity, uncertainty, and model–process mismatch.

The numerical record consists of reported manuscript summaries and embedded figures. The original experiment program, raw outputs and run logs were not available with the inspected manuscript; no independent regeneration of this campaign is claimed. The conclusions are conditional on the stated configurations and require the original materials for reproducibility verification.

The macro-state construction, archive assignment and empirical transition estimate limit the model to a small reported regime space. Rare transitions can be dominated by a prior, while continuous or changing regimes would require a separately specified classification policy. The campaign provides no measured observation-noise, false-alarm or simulator–plant mismatch study. Missing kinetic constants, initial states, decision bounds, random-stream rules and event counters also limit independent reproduction. Predictive understanding of community dynamics is a broader challenge [26], not an outcome established by this surrogate.

Robust design and model-predictive control are meaningful alternatives when the decision problem permits them [9], [32]. Neither is excluded by nutrient switching, and neither was benchmarked here. Evolutionary search may be useful for multimodal simulator objectives, but the present fixed-species continuous design does not establish a combinatorial community-assembly advantage. The restricted evidence supports comparison of the stated configurations rather than a general dominance claim.