PaperlayerView exact revision on arXiv

Source-linked AI summary

Inherited Learning in an Artificial Ecology: How Controls and Update Allocation Shape Benefits

Xuening Wu, Lei Li, Shan Yu

arXiv:2610.01232v1cs.NE

TL;DR

The paper asks when inherited learning helps populations retain useful experience and how to separate that benefit from control and update-allocation effects. In a resource-limited artificial ecology, it combines structured random controls with newborn and allocation interventions. State-preserving controls narrow the apparent advantage, while conditional speed benefits remain and allocation can change rankings.

  • Problem

    The paper asks how populations can retain useful experience across individual replacement and how inheritance benefits can be distinguished from effects of update structure, stability, and reproduction-linked allocation.

    Source: §1 · p. 1

  • Method

    The study uses a resource-limited artificial ecology with structured random controls and interventions on newborn preference expression and hereditary-update allocation.

    Source: §1 · p. 2, §3.4 · p. 6

  • Results

    State-preserving random controls narrow the apparent establishment advantage while retaining a conditional speed benefit, and update-allocation choices can change rankings.

    Source: §6 · p. 14

  • Takeaways & Limitations

    Evaluating inherited learning in this ecology requires accounting for update structure, newborn expression, and the coupling between reproduction and hereditary-update allocation.

    Source: §1 · p. 2, §6 · p. 14

    The benefit is conditional: no-writeback references went extinct in the two specified environments, but the reference established in 20/20 runs at low mortality.

    Source: §4.3 · p. 9, §4.9 · p. 13

Abstract

from arXiv · show

Learning can improve an individual's behavior, yet a population risks losing that experience whenever individuals die and are replaced. Inheriting learned preferences offers a way to preserve useful behavior across generations, raising a question for artificial populations: when does inheritance improve collective performance, and how can its benefits be measured fairly? The challenge is that inheritance changes not only offspring behavior but also survival, reproduction, and opportunities for further hereditary updates. Random controls with equal update magnitudes may therefore yield misleading comparisons if they alter different states or obey different stability constraints. We investigate this problem in a resource-limited artificial ecology, combining structured random controls with interventions on newborn preferences and the allocation of hereditary updates. Preserving the state structure of random updates substantially narrows the apparent inheritance advantage, while a conditional establishment-speed benefit remains. Preference erasure and faster-learning compensation support a contribution from reduced offspring relearning. Update allocation also changes the comparison: event quotas and common time cutoffs can reverse rankings, although they also change realized update amounts. With update count and cumulative magnitudes matched, staged release improves occupancy but does not achieve the prespecified establishment criterion. These findings provide a framework for distinguishing the value of inherited preferences from the effects of control design and update allocation, clarifying how inherited learning should be evaluated in artificial populations.

1 Introduction

The paper asks how inherited learning can preserve useful experience in a population while distinguishing inheritance benefits from effects of control design and update allocation. It studies these questions through interventions in a resource-limited artificial ecology.

1 Introduction

  • The central challenge is that inherited parameters may reduce offspring relearning while also changing survival, reproduction, and later opportunities for inheritance.The study therefore connects individual behavior to population dynamics.Source: §1 · p. 1
  • The study distinguishes the benefit of learned update direction from the experimental choices used to measure inheritance.Prior work has examined learning-mediated evolutionary change and inherited controllers, including newborn performance [2].Source: §1 · p. 1
  • Equal-norm random updates can differ in which states they affect, while directed and random updates can face different stability constraints.Because reproduction determines when hereditary updates occur, equal event quotas may also stop updates at different times across populations.Source: §1 · pp. 1–2
  • The study uses a resource-limited artificial ecology with local learning, stochastic death, reproduction, and inherited action-value tables.The external ecological rules remain fixed while local resources and population composition evolve; the testbed does not impose an external population-level fitness selection operator.Source: §1 · p. 2
  • State-preserving random controls sharply narrow the establishment-attainment gap at a fixed 512-event budget while retaining a speed advantage.The paper also finds that preference erasure impairs establishment, faster newborn learning compensates for much of that loss, and staged release improves occupancy without reaching the prespecified threshold.Source: §1 · p. 2
  • The contribution is a sequence of controls and interventions, supported by quota, representation, and ecological sensitivity analyses.The comparisons estimate distinct effects: matching candidate geometry does not match diverging histories, and improved occupancy is not evidence of faster establishment.Source: §1 · p. 2

2 Related work

The paper situates inherited learning among work on learning-mediated evolution, inherited robotic controllers, and evolutionary search allocation. It frames its contribution as a targeted comparison in an ecology with local resources and variable occupancy, not as a universal result.

Related work

  • Earlier work covers learning-mediated evolutionary change [2], inherited robotic controllers and newborn performance, and parental sample reuse in Bayesian optimization [7].These studies preclude treating reduced repeated learning as a new general finding.Source: §2 · p. 3
  • The paper distinguishes its comparison by preserving candidate update state-wise geometry and intervening on expressed newborn preferences while holding inherited parameters fixed after the quota.Its hereditary event quota is distinct from a controller evaluation budget.Source: §2 · p. 3
  • Inheritance disadvantages in stationary settings are not presented as novel, since analyses identify settings where Baldwinian evolution outperforms Lamarckian evolution [10].Here, lifetime learning continues after new hereditary writes stop.Source: §2 · p. 3
  • Resource-dependent rankings are not claimed as new; the paper contrasts its variable occupancy and local resource depletion with prior fixed-size, nonspatial populations [6].It also notes studies of environmental conflict, predictability, and reduced inheritance benefits under pressure for morphological novelty [8] [9], while making no claim of a common universal law.Source: §2 · p. 3

3 Model and experimental methods

The methods combine a spatial resource ecology and lifetime action-value learning with hereditary writeback, structured random controls, newborn interventions, and alternative update schedules. Outcomes are assessed with a prespecified establishment endpoint and paired seed comparisons.

3.1 Spatial ecology

  • The spatial ecology is a 32 × 32 toroidal lattice with replenishing local resources, stochastic mortality, and rest, harvest, and divide actions.At most one individual occupies each site; successful division requires sufficient energy and an empty neighboring site, with child energy share varied in specified experiments.Source: §3.1 · pp. 3–4

3.2 State, action selection, and lifetime learning

  • Individuals use inherited table g to initialize lifetime values θ across 128 energy-resource-feasibility states, with three action values per state.Actions are chosen by ε-greedy selection with ε = 0.05, and the selected value receives an immediate action-value update without a next-state bootstrap.Source: §3.2 · p. 4

3.3 Hereditary update and random controls

  • At birth, proportional writeback adds a fraction of the parent’s θ − g difference to g; central experiments set mutation to zero and compare against λ = 0 with lifetime learning retained.The no-writeback reference is not an independent-mutation evolutionary algorithm.Source: §3.3 · p. 4
  • The stronger random control preserves state-wise geometry by randomizing action preference within states, whereas weaker whole-table controls move updates across states.A quota limits nonzero hereditary updates, not births or learning interactions; equal fixed-norm update counts do not equalize native writeback magnitude totals.Source: §3.3 · p. 5

3.4 Newborn expression intervention

  • The newborn intervention branches after the 512th hereditary update, either keeping θ = g or replacing newborn expressed values with per-state means while preserving g.The erase-versus-keep contrast is within each direction at ηc = 0.2; ηc = 0.8 tests faster-learning compensation, but the branches do not provide a complete mediation decomposition.Source: §3.4 · pp. 5–6

3.5 Symmetric constraints and allocation diagnostics

  • Allocation diagnostics impose common norm and coordinate-range constraints, compare event quotas with shared time cutoffs, and test early versus staged release of 1,024 fixed-size updates.The staged schedule releases 128 events at eight specified times; its fixed step size is an intervention rather than a replication of native proportional magnitudes.Source: §3.5 · p. 6

3.6 Design, endpoints, and statistical unit

  • Establishment is the first sampled time followed by four more samples with occupancy at least 0.95; otherwise capped time is 20,000 sweeps.Runs last 20,000 sweeps, comparisons use 20 paired seeds with descriptive bootstrap intervals, and the synthesis is not one prospectively registered study.Source: §3.6 · pp. 6–7

4 Results

The results test how control geometry, newborn preferences, and update allocation shape the apparent benefit of directed inheritance. Across these comparisons, the ranking depends on experimental design, and improved occupancy does not always meet the establishment criterion.

4.1 Control geometry changes the apparent benefit

  • At B = 512, directed inheritance established in 20/20 runs versus 18/20 for state-preserving randomization, with mean capped times of 2,290 versus 7,160 sweeps.The paired random-minus-directed difference was 4,870 [3,040, 7,120] sweeps; the apparent gap was larger with whole-table shuffling and isotropic variation, whose structural differences do not isolate a cause.Source: §4.1 · pp. 7–8, Figure 2 · p. 8

4.2 Newborn expression contributes to establishment

  • Erasing newborn preferences reduced directed attainment from 20/20 to 6/20, while raising the erased branch’s learning rate to 0.8 restored attainment to 20/20.At ηc = 0, erased populations all became extinct; the intervention supports a contribution from reconstructing preferences but does not establish that this burden mediates the entire directed advantage.Source: §4.2 · p. 8, Figure 3 · p. 9

4.3 Continuous native writeback retains a benefit

  • Under continuous native writeback, directed inheritance was faster in every paired seed across six tested f × λ combinations and yielded higher normalized occupancy AUC.The comparison does not isolate direction at equal cumulative variation because histories produce different update magnitudes and counts.Source: §4.3 · p. 9

4.4 Stopping writeback can change the ranking

  • At f = 0.65, stopping writeback at B = 1024 reversed the continuous-rule ranking at λ = 0.5 and λ = 1, while neither group established at λ = 0.25.At baseline f = 0.5, λ = 0.25, B = 1024, directed attainment was 7/20 versus 20/20, but directed AUC was slightly higher; random updating was not better on every metric.Source: Figure 4 · p. 10, §4.4 · p. 10

4.5 Benefits survive common stability constraints; 4.6 The event cutoff is endogenous to reproduction

  • Under common stability bounds, directed continuous updates remained faster across tested conditions, but bounded event quotas could reverse rankings as populations exhausted quotas at different times.The common-time contrast also changed realized write amounts: the 750-sweep condition used approximately 3,910 directed writes versus 1,094 random writes, so it was not a pure timing comparison at equal update amounts.Source: §4.5 · pp. 10–11, §4.6 · p. 11

4.7 Matched-budget staging improves occupancy, not establishment

  • With update count and cumulative magnitudes matched, staged release improved occupancy but none of the 80 runs reached the primary establishment threshold.Staging improved directed AUC by 0.08249 [0.07842,0.08664], while the prespecified faster-establishment prediction was not supported.Source: §4.7 · pp. 11–12, Table 5 · p. 12

4.8 Quota and representation sensitivity; 4.9 Ecological and numerical limits

  • Quota and representation checks retained directed speed ordering in several tested settings, but ecological sensitivities included environments where inheritance was unnecessary or insufficient for establishment.At low mortality, the no-writeback reference established in 20/20 runs; under severe mortality, both directions became extinct before using the quota. A high-offspring-share reversal depended on update rule and cutoff, not offspring share alone.Source: §4.8 · p. 12, §4.9 · p. 13

5 Discussion

The discussion finds a conditional establishment-speed benefit whose interpretation depends on the control, newborn expression, and allocation of hereditary updates. It also limits transfer to other representations and ecological settings.

5 Discussion

  • State-preserving controls narrow the apparent inheritance advantage, but preserve some learning information and do not match changes in action probabilities.Extending the control to other representations requires preserving the structure relevant to their behavior.Source: §5 · p. 13
  • Preference erasure and faster newborn learning support reduced relearning as a contributor, but do not establish complete mediation or identify the mechanism of native proportional writeback.The interventions retain inherited tables and row means and differ from the quota-exhausting birth onward.Source: §5 · p. 13
  • Event quotas can end hereditary updates at different times across populations, while common-time cutoffs change the coupling but also change realized update amounts.Matching update counts and cumulative magnitudes supports an allocation-policy effect on occupancy, but timing, state coverage, geometric feasibility, and selection remain coupled.Source: §5 · pp. 13–14
  • Transfer is limited by the simple ecology, immediate-value learner, engineered reward, explicit division-feasibility input, and absence of independent mutation in core comparisons.Establishment is sampled only at set times, each formal condition has 20 seeds, and the evidence does not establish increased long-term evolvability.Source: §5 · p. 14

6 Conclusion

In this resource-limited artificial ecology, inherited preferences have conditional population-level benefits, and the measured value depends on how preferences are expressed, transmitted, and allocated. The findings support a framework for evaluating inherited learning while limiting claims to this ecology.

6 Conclusion

  • State-preserving random controls reduce a large apparent establishment advantage to a conditional speed benefit, while newborn preference erasure and faster-learning compensation support reduced relearning as a contributor.Continuous proportional inheritance retains a speed advantage under common stability constraints.Source: §6 · p. 14
  • Event quotas and shared time cutoffs can produce different rankings while changing realized update amounts.When update count and cumulative magnitudes are matched, staged release improves time-averaged occupancy, but no group reaches the prespecified establishment threshold.Source: §6 · p. 14
  • The framework calls for accounting for update structure and stability, newborn expression, and reproduction’s coupling to update allocation when evaluating inherited learning.Testing whether these distinctions predict outcomes in an independently implemented learning system is identified as a next step toward broader applicability.Source: §6 · p. 14

Author declarations

The authors report no funding or competing interests and state that the experiments involved computational simulations without human participants, animals, or biological cell experiments.

Author declarations

A Additional ecological outcomes

The ecological sensitivity analysis varies mortality, offspring share, and birth bonus around a baseline, with paired directed and random outcomes reported for groups of 20 seeds.

A Additional ecological outcomes

  • Six fixed-norm sensitivity settings at B = 1024 each change one baseline setting: mortality, offspring share, or birth bonus.The baseline is mortality 0.002, offspring share 0.5, and birth bonus 0.2; alternatives set mortality to 0.0005 or 0.01, offspring share to 0.35 or 0.65, or bonus to zero.Source: A Additional ecological outcomes · p. 15
  • Table 7 presents paired directed and random entries for groups of 20 seeds, with Reached and Extinct reported as separate counts out of 20.The high-mortality capped-time tie reflects nonattainment in both groups, not equivalence.Source: Table 7 · p. 15

B Native magnitude and numerical audit

Native random updates accumulated larger magnitudes than directed inheritance, and their different stability constraints limit interpretation of the comparison. A bounded follow-up rules out unbounded random drift as necessary for a continuous benefit, but does not equalize cumulative changes or decompose the original effect.

B Native magnitude and numerical audit

  • For f = 0.5, λ = 0.5, and B = 1024, directed/random mean increment-norm sums were 99.10/158.09 and squared-norm sums were 13.28/30.03.Under continuous writeback at f = 0.65 and λ = 1, the random squared-sum median was approximately 15,343 versus 567 for directed inheritance.Source: B Native magnitude and numerical audit · p. 15
  • The random group’s mean squared increment sum was approximately 480,899, dominated by a retained run with squared sum 9.31 × 10^6 and maximum step norm approximately 1,664.48.A diagnostic replay reproduced saved outputs exactly, but numerical validity does not remove the scientific limitation.Source: B Native magnitude and numerical audit · p. 15
  • Directed native inheritance is a coordinate-wise convex combination, whereas state rotation preserves row geometry but not that convex-hull property.Differential stability and variation magnitude limit interpretation; the bounded follow-up rules out unbounded random drift as necessary for a continuous benefit, but does not decompose the original effect or equalize cumulative changes.Source: B Native magnitude and numerical audit · p. 15

C Reproducibility and provenance

The study records simulation provenance and checks implementation, population traces, and update ledgers. These checks document validation performed but do not constitute independent verification or reimplementation.

C Reproducibility and provenance

  • The authors state that AI assistance supported simulation, analysis, literature discovery, figures, and manuscript preparation, and that it was not independent verification.The 1,440 subsequent formal runs received corresponding checks, and new seed blocks were not extended after outcome inspection.Source: C Reproducibility and provenance · pp. 15–16
  • C++17 simulations used separate random streams, per-stage seeds and source snapshots, and recorded source and protocol hashes.Matched seeds align initial conditions and stream initialization, but do not enforce identical encounters after trajectories diverge.Source: C Reproducibility and provenance · p. 16
  • Checks covered population accounting, resource bounds, energy conservation, geometry, zero-learning and zero-writeback comparisons, and exact replay of representative outputs.The zero-learning check compares writeback variants within the same learning-rate setting, not learning against no learning without writeback.Source: C Reproducibility and provenance · p. 16
  • A large random-update replay used scale-relative tolerances and reproduced seven archived output types byte for byte, while its maximum reported absolute discrepancy was 4.66 × 10^-10.An initially incorrect implementation-check comparison was corrected before the native formal batch.Source: C Reproducibility and provenance · p. 16
  • All 1,840 formal population traces were reread to recompute attainment, capped times, occupancy, and extinction, matching archived seed summaries.This was a data-to-manuscript consistency check, not an independent reimplementation or new simulation experiment.Source: C Reproducibility and provenance · p. 16

D Scope of the literature comparison

The related-work assessment draws on targeted searches and primary texts available through 30 September 2026. It is explicitly neither a systematic review nor support for a first-ever claim.

D Scope of the literature comparison

  • The related-work assessment used targeted searches and primary texts available through 30 September 2026; it was not a systematic review and does not substantiate a first-ever claim.Closely related effects and mechanisms are acknowledged rather than recast as newly discovered principles.Source: D Scope of the literature comparison · p. 16
  • The 2023 robot preprint and journal publication are treated as one work, while two other 2026 robot works are cited as preprints without claims of final conference acceptance.The Bayesian-optimization sample-inheritance work is cited by its verified 2026 arXiv version without claiming that upload was its earliest dissemination.Source: D Scope of the literature comparison · p. 16
  • The literature comparison avoids asserting that the surveyed related effects or mechanisms are newly discovered principles.The assessment does not substantiate a first-ever claim.Source: D Scope of the literature comparison · p. 16
Loading 2610.01232v1…