Reference library

Provably Sub-Linear Two-Timescale NeuroEvolution with Online Plasticity

Paper summary

Separates architecture evolution from reward-modulated Hebb/Oja/BCM weight adaptation. Authors report gains over fixed-weight NEAT on three continuous-control tasks, while PPO/SAC remain stronger on Hopper and BipedalWalker under reported interaction budgets. A simplified regret bound assumes bounded projected parameters, a finite architecture pool, full-information Hedge selection and surrogate domination. Appendix C.9 reports similar outcomes with and without inherited adapted weights.

Relevance to the project

Assistant interpretation: separate the supplied lifetime reward, inherited rule family, genotype parameters and acquired state when comparing adaptive mechanisms. Neither reported similarity of inheritance variants nor the conditional theorem chooses our F/H inheritance contract or supplies endogenous credit assignment.

Synaptic plasticityLearning and adaptationGenetic representationsReinforcement learning

The theorem is conditional on A5, not implied by reward-gated local activity. Simplified full-information selection omits practical NEAT growth/speciation. Budget accounting uses rollout length and does not establish matched total compute including tuning. Similar inheritance curves are an author report, not an equivalence proof.

Source record

Reference ID: arxiv-2606-20817
Archived version: arXiv 2606.20817v1; version-specific PDF URL; intended title/authors checked in first page
Summary provenance: reference-library/papers/arxiv-2606-20817/metadata.json · summary
Review status: targeted full-text methods/results reading; not an exhaustive critical review
Reading scope: Identity/abstract; selected PDF pp.3–7 (sections3.2–6, assumptions A1–A7, main regret statement, methods/budgets, results, ablation and limitations), appendix C.9 text on p.22 and opening proof passages on p.11. Primary HTML cross-check of equations/assumptions and C.9. Figures read through captions/text, not a visual audit. No exhaustive appendix proof review, code inspection/execution, raw-result verification or replication.