pith. sign in

arxiv: 2604.12336 · v1 · submitted 2026-04-14 · 💻 cs.NE · cs.AI

GeM-EA: A Generative and Meta-learning Enhanced Evolutionary Algorithm for Streaming Data-Driven Optimization

Yue Wu , Yuan-Ting Zhong , Ze-Yuan Ma , Yue-Jiao Gong This is my paper

Pith reviewed 2026-05-10 14:29 UTC · model grok-4.3

classification 💻 cs.NE cs.AI
keywords streaming data-driven optimizationevolutionary algorithmsmeta-learninggenerative replayconcept driftsurrogate adaptationnon-stationary environments
0
0 comments X

The pith

GeM-EA uses bi-level meta-learning to initialize surrogates from historical priors and generative replay to leverage past solutions, enabling faster adaptation to concept drift in streaming data-driven optimization.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper aims to show that existing methods for streaming data-driven optimization struggle with concept drift because they either combine surrogates simply or inject old solutions directly, often leading to poor performance. GeM-EA addresses this by detecting drifts and applying a bi-level meta-learning strategy for quick surrogate setup plus generative replay in multi-island evolution. If successful, this would allow optimization methods to maintain effectiveness in continuously changing environments without frequent restarts or negative effects from outdated information. Readers interested in real-time decision systems or adaptive machine learning would find this relevant as it targets a common challenge in dynamic settings.

Core claim

GeM-EA unifies meta-learned surrogate adaptation with generative replay for effective evolutionary search in SDDO. Upon detecting concept drift, a bi-level meta-learning strategy rapidly initializes the surrogate using environment-relevant priors, while a linear residual component captures global trends. A multi-island evolutionary strategy further leverages historical knowledge via generative replay to accelerate optimization, demonstrating faster adaptation and improved robustness on benchmark problems compared with state-of-the-art methods.

What carries the argument

The unification of bi-level meta-learning for surrogate adaptation and generative replay within a multi-island evolutionary strategy.

Load-bearing premise

The approach depends on accurate concept drift detection and the availability of relevant historical environments for meta-learning, as mismatches could lead to ineffective initialization.

What would settle it

A direct test would involve running GeM-EA on SDDO benchmarks modified to include sudden drifts that evade detection or use irrelevant historical data, checking if the adaptation speed and robustness advantages disappear.

Figures

Figures reproduced from arXiv: 2604.12336 by Yuan-Ting Zhong, Yue-Jiao Gong, Yue Wu, Ze-Yuan Ma.

Figure 1
Figure 1. Figure 1: The framework of GeM-EA confidence-driven migration mechanism coordinates their exchange, isolating the search from misleading historical biases. • Extensive experiments demonstrate that GeM-EA consis￾tently achieves superior solution quality, faster adaptation, and high computational efficiency compared with state-of￾the-art SDDO methods. 2 Methodology As illustrated in [PITH_FULL_IMAGE:figures/full_fig_… view at source ↗
Figure 2
Figure 2. Figure 2: Online Error convergence trajectories on the last [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
read the original abstract

Streaming Data-Driven Optimization (SDDO) problems arise in many applications where data arrive continuously and the optimization environment evolves over time. Concept drift produces non-stationary landscapes, making optimization methods challenging due to outdated models. Existing approaches often rely on simple surrogate combinations or directly injecting solutions, which may cause negative transfer under sudden environmental changes. We propose GeM-EA, a Generative and Meta-learning Enhanced Evolutionary Algorithm for SDDO that unifies meta-learned surrogate adaptation with generative replay for effective evolutionary search. Upon detecting concept drift, a bi-level meta-learning strategy rapidly initializes the surrogate using environment-relevant priors, while a linear residual component captures global trends. A multi-island evolutionary strategy further leverages historical knowledge via generative replay to accelerate optimization. Experimental results on benchmark SDDO problems demonstrate that GeM-EA achieves faster adaptation and improved robustness compared with state-of-the-art methods.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 0 minor

Summary. The paper proposes GeM-EA, a generative and meta-learning enhanced evolutionary algorithm for streaming data-driven optimization (SDDO) under concept drift. It combines bi-level meta-learning to rapidly initialize surrogates from environment-relevant priors upon drift detection, a linear residual component to capture global trends, and a multi-island evolutionary strategy that uses generative replay to leverage historical knowledge. The central claim is that this unified approach yields faster adaptation and improved robustness relative to state-of-the-art methods on benchmark SDDO problems.

Significance. If the empirical claims hold under rigorous validation, the work offers a concrete mechanism for mitigating negative transfer in non-stationary surrogate-assisted optimization by coupling meta-learned initialization with generative replay. This integration could influence algorithm design for dynamic real-world tasks such as online hyperparameter tuning or adaptive control systems. The paper's contribution is the explicit unification of these components rather than incremental improvements to any single technique.

major comments (2)
  1. [Abstract / Experimental Results] Abstract and Experimental Results section: the claim that GeM-EA 'achieves faster adaptation and improved robustness' rests on benchmark results whose details (specific SDDO test problems, drift types and frequencies, number of runs, performance metrics, statistical tests, and exact baselines) are not provided, rendering the primary empirical support unverifiable and load-bearing for the paper's conclusion.
  2. [Method / §3] Method description (bi-level meta-learning and generative replay components): no explicit stress test or ablation is reported for the case in which detected drifts produce environments whose relevant priors differ substantially from the meta-training distribution; this leaves the acknowledged negative-transfer risk from prior work unaddressed in the evaluation.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback. We address each major comment point by point below and describe the revisions we will incorporate to improve the manuscript.

read point-by-point responses

  1. Referee: [Abstract / Experimental Results] Abstract and Experimental Results section: the claim that GeM-EA 'achieves faster adaptation and improved robustness' rests on benchmark results whose details (specific SDDO test problems, drift types and frequencies, number of runs, performance metrics, statistical tests, and exact baselines) are not provided, rendering the primary empirical support unverifiable and load-bearing for the paper's conclusion.

    Authors: We agree that the experimental claims require more explicit and structured details to be fully verifiable. While the Experimental Results section presents benchmark comparisons, we acknowledge that a consolidated summary of the test problems, drift parameters, run counts, metrics, baselines, and statistical analysis would strengthen transparency. In the revised manuscript we will add a table and accompanying text in the Experimental Results section that lists the specific SDDO benchmark problems (including function names, dimensions, and drift characteristics such as sudden/gradual drift types and frequencies), the number of independent runs (30), the performance metrics (mean best fitness and adaptation speed), the exact baseline algorithms, and the results of statistical significance tests (Wilcoxon rank-sum tests with p-values). The abstract claim will remain unchanged but will be explicitly tied to these details. This constitutes a clear revision. revision: yes

  2. Referee: [Method / §3] Method description (bi-level meta-learning and generative replay components): no explicit stress test or ablation is reported for the case in which detected drifts produce environments whose relevant priors differ substantially from the meta-training distribution; this leaves the acknowledged negative-transfer risk from prior work unaddressed in the evaluation.

    Authors: We accept that an explicit stress test for substantial mismatch between meta-training priors and post-drift environments would more directly address the negative-transfer concern. Our existing experiments cover a variety of drift scenarios on standard benchmarks, but they do not isolate the extreme mismatch case. In the revised version we will add a dedicated ablation subsection that introduces controlled environments with priors deliberately distant from the meta-training distribution (e.g., using held-out or synthetically shifted environment parameters). We will report performance of the full GeM-EA, the version without meta-learning, and the version without generative replay, thereby quantifying how the bi-level initialization and replay components mitigate negative transfer. This addition will be included in the Method and Experimental Results sections. revision: yes

Circularity Check

0 steps flagged

No significant circularity in derivation chain

full rationale

The paper describes an algorithmic framework (GeM-EA) that combines concept drift detection, bi-level meta-learning for surrogate initialization, a linear residual component, and multi-island generative replay. No equations, first-principles derivations, or predictions are presented that reduce by construction to fitted parameters, self-definitions, or self-citations. The central empirical claim rests on benchmark experiments comparing adaptation speed and robustness, which are independent of the method's internal structure and do not invoke load-bearing self-citations or uniqueness theorems. The approach is self-contained as a proposed heuristic without circular reduction.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract provides no technical details on specific parameters or assumptions; the approach appears to build on standard concepts in meta-learning and evolutionary algorithms without introducing new entities or ad-hoc axioms visible here.

pith-pipeline@v0.9.0 · 5458 in / 1130 out tokens · 28831 ms · 2026-05-10T14:29:16.769969+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

22 extracted references · 22 canonical work pages

  1. [1]

    Xu Chen, Junshan Wang, and Kunqing Xie. 2021. TrafficStream: A streaming traffic flow forecasting framework based on graph neural networks and continual learning.IJCAI(2021)

    work page 2021

  2. [2]

    Chelsea Finn, Pieter Abbeel, and Sergey Levine. 2017. Model-agnostic meta- learning for fast adaptation of deep networks. InInternational conference on machine learning. PMLR, 1126–1135

    work page 2017

  3. [3]

    Chelsea Finn, Aravind Rajeswaran, Sham Kakade, and Sergey Levine. 2019. Online meta-learning. InInternational conference on machine learning. PMLR, 1920–1930

    work page 2019

  4. [4]

    Yue-Jiao Gong, Yuan-Ting Zhong, and Hao-Gan Huang. 2023. Offline data-driven optimization at scale: A cooperative coevolutionary approach.IEEE Transactions on Evolutionary Computation28, 6 (2023), 1809–1823

    work page 2023

  5. [5]

    Hongshu Guo, Zeyuan Ma, Yining Ma, Xinglin Zhang, Wei-Neng Chen, and Yue-Jiao Gong. 2025. DesignX: Human-Competitive Algorithm Designer for Black-Box Optimization.arXiv preprint arXiv:2505.17866(2025)

    work page arXiv 2025

  6. [6]

    Pengfei Huang, Handing Wang, and Yaochu Jin. 2021. Offline data-driven evolu- tionary optimization based on tri-training.Swarm and evolutionary computation 60 (2021), 100800

    work page 2021

  7. [7]

    Jian-Yu Li, Zhi-Hui Zhan, Chuan Wang, Hu Jin, and Jun Zhang. 2020. Boost- ing data-driven evolutionary algorithm with localized data generation.IEEE Transactions on Evolutionary Computation24, 5 (2020), 923–937

    work page 2020

  8. [8]

    Ke Li, Renzhi Chen, and Xin Yao. 2023. A data-driven evolutionary transfer optimization for expensive problems in dynamic environments.IEEE Transactions on Evolutionary Computation(2023)

    work page 2023

  9. [9]

    Wenjian Luo, Ruikang Yi, Bin Yang, and Peilan Xu. 2018. Surrogate-assisted evolutionary framework for data-driven dynamic optimization.IEEE Transactions on Emerging Topics in Computational Intelligence3, 2 (2018), 137–150

    work page 2018

  10. [10]

    Zeyuan Ma, Yue-Jiao Gong, Hongshu Guo, Wenjie Qiu, Sijie Ma, Hongqiao Lian, Jiajun Zhan, Kaixu Chen, Chen Wang, Zhiyang Huang, Zechuan Huang, Guojun Peng, Ran Cheng, and Yining Ma. 2025. MetaBox-v2: A Unified Benchmark Platform for Meta-Black-Box Optimization. InAdvances in Neural Information Processing Systems

    work page 2025

  11. [11]

    Ewa Osekowska, Henric Johnson, and Bengt Carlsson. 2017. Maritime vessel traffic modeling in the context of concept drift.Transportation research procedia 25 (2017), 1457–1476

    work page 2017

  12. [12]

    Jooyoung Park and Irwin W Sandberg. 1991. Universal approximation using radial-basis-function networks.Neural computation3, 2 (1991), 246–257

    work page 1991

  13. [13]

    Maycon LM Peixoto, Edson Mota, Adriano HO Maia, Wellington Lobato, Mo- hammad A Salahuddin, Raouf Boutaba, and Leandro A Villas. 2023. FogJam: a fog service for detecting traffic congestion in a continuous data stream VANET. Ad Hoc Networks140 (2023), 103046

    work page 2023

  14. [14]

    Wen-Jie Qiu, Hong-Shu Guo, Ze-Yuan Ma, Jun Zhang, and Yue-Jiao Gong. 2026. Automated Algorithm Design for Black-Box Optimization: Progress and Chal- lenges.Chinese Journal of Computers49, 4 (April 2026). doi:10.11897/SP.J.1016. 2026.00855

    work page doi:10.11897/sp.j.1016 2026

  15. [15]

    Sachin Ravi and Hugo Larochelle. 2017. Optimization as a model for few-shot learning. InInternational conference on learning representations

    work page 2017

  16. [16]

    Oriol Vinyals, Charles Blundell, Timothy Lillicrap, Daan Wierstra, et al . 2016. Matching networks for one shot learning.Advances in neural information pro- cessing systems29 (2016)

    work page 2016

  17. [17]

    Gerhard Widmer and Miroslav Kubat. 1996. Learning in the presence of concept drift and hidden contexts.Machine learning23, 1 (1996), 69–101

    work page 1996

  18. [18]

    Cuie Yang, Jinliang Ding, Yaochu Jin, and Tianyou Chai. 2023. A data stream ensemble assisted multifactorial evolutionary algorithm for offline data-driven dynamic optimization.Evolutionary Computation31, 4 (2023), 433–458

    work page 2023

  19. [19]

    Huan Zhang, Jinliang Ding, Liang Feng, Kay Chen Tan, and Ke Li. 2024. Solving expensive optimization problems in dynamic environments with meta-learning. IEEE transactions on cybernetics(2024)

    work page 2024

  20. [20]

    Yuanting Zhong, Xincan Wang, Yuhong Sun, and Yue-Jiao Gong. 2024. Sd- dobench: A benchmark for streaming data-driven optimization with concept drift. InProceedings of the Genetic and Evolutionary Computation Conference. 59–67

    work page 2024

  21. [21]

    Yuan-Ting Zhong and Yue-Jiao Gong. 2025. Data-Driven Evolutionary Compu- tation Under Continuously Streaming Environments: A Drift-Aware Approach. IEEE Transactions on Evolutionary Computation(2025)

    work page 2025

  22. [22]

    Yuan-Ting Zhong, Ting Huang, Xiaolin Xiao, and Yue-Jiao Gong. 2026. TRACE: A Generalizable Drift Detector for Streaming Data-Driven Optimization. InPro- ceedings of the AAAI Conference on Artificial Intelligence, Vol. 40. 28910–28918

    work page 2026