Paper deep dive
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
Intelligence
Status: succeeded | Model: google/gemini-3.1-flash-lite-preview | Prompt: intel-v1 | Confidence: 98%
Last extracted: 4/15/2026, 1:36:56 AM
Summary
GeM-EA is a novel evolutionary algorithm designed for Streaming Data-Driven Optimization (SDDO) that addresses concept drift by combining bi-level meta-learning for surrogate adaptation with a multi-island generative replay strategy. It decouples structural and non-structural parameter optimization to ensure stability and uses a confidence-driven migration mechanism to prevent negative transfer, achieving superior convergence and linear computational complexity.
Entities (5)
Relation Signals (4)
Concept Drift → affects → SDDO
confidence 100% · Concept drift produces non-stationary landscapes, making optimization methods challenging
GeM-EA → evaluatedon → SDDObench
confidence 100% · We utilize SDDObench (Zhong et al., 2024) .
GeM-EA → solves → SDDO
confidence 100% · We propose GeM-EA, a Generative and Meta-learning Enhanced Evolutionary Algorithm for SDDO
GeM-EA → uses → RBFN
confidence 95% · We employ a Radial Basis Function Network (RBFN) as the base surrogate
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
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 Concept drift produces non-stationary landscapes, making optimization methods challenging due to outdated models.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.
Tags
Links
- Source: https://arxiv.org/abs/2604.12336v1
- Canonical: https://arxiv.org/abs/2604.12336v1
Trouble viewing inline? Open PDF directly →
Full Text
28,962 characters extracted from source content.
Expand or collapse full text
by 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† South China University of TechnologyGuangzhouChina ∗Equal contribution. †Corresponding author: gongyuejiao@gmail.com (2026) 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.The source code is available at https://github.com/PoetMoon/GeM-EA. Streaming data-driven evolutionary algorithms, Data stream, Meta-learning, Concept drift †journalyear: 2026†copyright: c†conference: Genetic and Evolutionary Computation Conference; July 13–17, 2026; San Jose, Costa Rica†booktitle: Genetic and Evolutionary Computation Conference (GECCO Companion ’26), July 13–17, 2026, San Jose, Costa Rica†doi: 10.1145/3795101.3805285†isbn: 979-8-4007-2488-6/2026/07†ccs: Theory of computation Evolutionary algorithms†ccs: Information systems Data streams 1. Introduction With ongoing technological advances, applications like smart city management (Chen et al., 2021; Peixoto et al., 2023) and intelligent transportation systems (Osekowska et al., 2017) increasingly rely on continuous data streams. Optimization in such settings is inherently dynamic, as the underlying data distribution evolves over time—a phenomenon known as concept drift (Widmer and Kubat, 1996). This context defines Streaming Data-Driven Optimization (SDDO) problems (Zhong et al., 2024), requiring methods to adapt dynamically to maintain performance. To address SDDO, Streaming Data-Driven Evolutionary Algorithms (SDDEAs) integrate Evolutionary Algorithms (EAs) with data-driven modeling (Gong et al., 2023). While existing methods, such as explicit historical solution retrieval (Zhong and Gong, 2025), attempt to exploit past knowledge, current approaches still suffer from rigid knowledge reuse. Specifically, linear surrogate combinations fail to capture complex geometric transformations like rotation. Such spatial shifts introduce severe asymmetry and non-separability into the optimization landscape. Consequently, injecting individuals based on these misaligned models creates a high risk of negative transfer, misleading the search under sudden drifts. Meta-learning (Finn et al., 2017) offers a promising “learning to adapt” paradigm that transcends linear ensembles by learning malleable initializations for complex landscape changes. Despite success in general machine learning (Ravi and Larochelle, 2017; Vinyals et al., 2016), meta-learning remains largely unexplored in SDDO (Zhang et al., 2024). Existing attempts require active data sampling, a practice strictly prohibited in SDDO. Furthermore, traditional meta-learning assumes tasks are sampled from a stationary distribution (Finn et al., 2019) and relies on discrete task formulations. These assumptions fundamentally clash with SDDO’s continuous streaming nature, where concept drift is unpredictable and explicit environment segmentation is unavailable. To overcome these limitations, we propose GeM-EA, a Generative and Meta-learning Enhanced Evolutionary Algorithm for SDDO. Our main contributions are: Figure 1. The framework of GeM-EA • A unified framework synergizing meta-learned surrogate adaptation with generative replay-based evolutionary search. It rapidly initializes the surrogate upon concept drift (Zhong et al., 2026) without active data acquisition and distills historical solutions via generative replay, accelerating search and mitigating negative transfer. • A bi-level meta-learned surrogate adaptation module. It decouples structural topology adaptation from fast adaptation by meta-optimizing structural parameters while analytically solving non-structural widths and weights. This alleviates instability in fully gradient-based meta-learning, while a linear residual component further captures global landscape trends. • A generative replay-based evolutionary search utilizing a multi-island architecture. A meta-adaptation island aggressively explores the current landscape, while history-anchor islands maintain robustness using generated historical solutions. A confidence-driven migration mechanism coordinates their exchange, isolating the search from misleading historical biases. • Extensive experiments demonstrate that GeM-EA consistently achieves superior solution quality, faster adaptation, and high computational efficiency compared with state-of-the-art SDDO methods. 2. Methodology As illustrated in Figure 1, GeM-EA synergizes meta-learned surrogate adaptation with generative replay-based evolutionary search. We employ a Radial Basis Function Network (RBFN) (Park and Sandberg, 1991) as the base surrogate, uniquely characterized by parameter separability into structural parameters (θ) and non-structural parameters (w,σw,σ). 2.1. Meta-Learned Surrogate Adaptation Directly optimizing the parameter set Ψ=θ,w,σ =\θ,w,σ\ in a unified meta-learning loop is computationally unstable due to severe gradient conflicts. This stems from: 1) convexity mismatch, as weights w follow a convex least-squares objective while structural parameters θ occupy a highly non-convex landscape; and 2) sensitivity disparity, since minor θ perturbations fundamentally reshape basis functions, whereas w changes only induce linear scaling. Consequently, joint updates frequently trigger pronounced oscillations or premature stagnation. To overcome this, we propose a bi-level meta-learned surrogate adaptation strategy that decouples structural parameter optimization from fast adaptation, enabling stable surrogate initialization and reliable meta-adaptation under streaming environments. Stage 1: Environment-Relevant Prior Identification. To mitigate negative transfer in streaming environments, we construct a meta-dataset metaS_meta by retrieving only historical environments highly relevant to the current dataset DtD_t. Relevance is evaluated via a discrepancy metric combining approximation error and prediction divergence: (1) Dist(Dt,ℳi)=γ1MAPE(Dt,ℳi)+γ21N∑∈Dt‖ℳtemp()−ℳi()‖2Dist(D_t,M_i)= _1MAPE(D_t,M_i)+ _2 1N _x∈ D_t||M_temp(x)-M_i(x)||^2 where MAPE denotes mean absolute percentage error, and ℳtempM_temp is a temporary surrogate trained exclusively on DtD_t serving as a structural baseline. Stage 2: Bi-Level Meta-Learned Fast Adaptation. Environment-level structural basis parameters θ update via gradient descent for topological shifts (θt′=θ−α∇θℒDtθ _t=θ-α _θL_D_t), and widths σt′σ _t recalibrate via k-nearest neighbor heuristics. Output weights wt′w _t are analytically resolved via Ridge Regression using the updated basis activation matrix Φ and ground-truth yty_t: (2) wt′=(Φ(θt′,σt′)TΦ(θt′,σt′)+λ)−1Φ(θt′,σt′)Tytw _t= ( (θ _t,σ _t)^T (θ _t,σ _t)+ )^-1 (θ _t,σ _t)^Ty_t Meta-level θ is gradient-refined minimizing cumulative generalization loss across metaS_meta (θ←θ−β∇θ∑DtℒDtθ←θ-β _θ _D_tL_D_t), decoupling structural tuning from analytical weight solving. Stage 3: Residual-Augmented Stabilization. To mitigate limited-data residual approximation errors, the final surrogate augments the meta-learned f^meta() f_meta(x) with a global linear residual: F^()=f^meta()+T+b F(x)= f_meta(x)+a^Tx+b. Linear parameters ,b\a,b\ are analytically solved via pseudo-inverse to minimize the residual Mean Squared Error (ej=yj−f^meta(j)e_j=y_j- f_meta(x_j)), efficiently balancing fast environment-specific adaptation and global stability. 2.2. Generative Replay-based Evolutionary Search Instead of relying on a single population vulnerable to negative transfer, GeM-EA employs a multi-island architecture. The meta-adaptation island conducts aggressive exploration uniformly guided by the meta-learned surrogate. Simultaneously, P history-anchor islands are generated from the statistical summaries of the top-P most similar past environments, acting as stabilizing anchors to accelerate exploitation. Algorithm 1 Generative Replay-based Evolutionary Search 0: History Archive ArcArc, Migration interval τ, Max generations GmaxG_max, Meta-learned surrogate F^() F(x) 0: Optimal solution meta∗x^*_meta 1: Initialization: 2: Select top-P metaS_meta from ArcArc 3: Randomly initialize meta-island Λmeta _meta 4: Generate anchor islands Λanchor(i)i=1P\ _anchor^(i)\_i=1^P based on metaS_meta 5: Evolutionary Search: 6: for g=1g=1 to GmaxG_max do 7: // Meta-Adaptation Island 8: Λmeta←Evolve(Λmeta,F^meta) _meta ( _meta, F_meta) 9: meta∗←Best(Λmeta)x^*_meta ( _meta) 10: // History-Anchor Islands 11: for i=1i=1 to P do 12: Λanchor(i)←Evolve(Λanchor(i),F^i) _anchor^(i) ( _anchor^(i), F_i) 13: i∗←Best(Λanchor(i))x^*_i ( _anchor^(i)) 14: end for 15: // Bidirectional Migration 16: if gmodτ=0g τ=0 then 17: Λmeta←Λmeta∪1∗,…,P∗ _meta← _meta∪\x^*_1,…,x^*_P\ 18: for i=1i=1 to P do 19: if F^i(meta∗)<F^i(i∗) F_i(x^*_meta)< F_i(x^*_i) then 20: Λanchor(i)←Λanchor(i)∪meta∗ _anchor^(i)← _anchor^(i)∪\x^*_meta\ 21: end if 22: end for 23: end if 24: end for 25: Update Archive ArcArc based on clustering 26: return meta∗x^*_meta To coordinate these islands, we implement a confidence-driven migration mechanism. At regular intervals τ, elite solutions from history-anchor islands are injected into the meta-adaptation island. Conversely, if the meta-adaptation island discovers superior basins of attraction, its best solution meta∗x^*_meta is fed back to an anchor island i only if it satisfies a strict confidence condition: F^i(meta∗)<F^i(i∗) F_i(x^*_meta)< F_i(x^*_i). This bidirectional exchange balances exploration and historical exploitation while effectively shielding the search from negative transfer. 3. Experimental Results 3.1. Experimental Settings We utilize SDDObench (Zhong et al., 2024) . Algorithm performance is evaluated using two standard metrics: offline error (EofflineE_offline) and online error(EonlineE_online). Detailed mathematical definitions of these metrics, along with extended related work on SDDO, are provided in the Supplementary Appendix 1. To ensure strictly fair comparisons, all evaluated algorithms (T-DDEA (Huang et al., 2021), BDDEA-LDG (Li et al., 2020), MLO (Zhang et al., 2024), DETO (Li et al., 2023), DSE-MFS (Yang et al., 2023), SAEF-1GP (Luo et al., 2018), and DASE (Zhong and Gong, 2025)) share the same evolutionary budget: a population size of 300 and a maximum of 50 iterations per environment. For GeM-EA, the RBFN uses Kc=⌊Ndata⌋K_c= N_data centers, λ=0.01λ=0.01, and a meta-learning rate of 10−410^-4. The DE optimizer uses DE/current-to-best/1 (F=0.5,Cr=0.9F=0.5,Cr=0.9) with P=4P=4 history-anchor islands and a migration interval τ=10τ=10. To ensure statistical robustness, each experiment is conducted 10 times independently; performance is reported as the mean ± standard deviation calculated from these repetitions. Statistical significance is rigorously assessed using the Kruskal-Wallis test followed by the post hoc Dunnett’s test with a Bonferroni correction strategy. The significance level is 0.05. To clearly visualize the comparative performance in the result tables, we use ”++” to denote that our method significantly outperforms the competitor, ”≈” to indicate statistically similar performance, and ”−-” to denote underperformance. 3.2. Comparative Analysis with State-of-the-Art Methods 3.2.1. Comparative Solution Quality Analysis. As shown in Table 1, GeM-EA achieves the lowest average rank (1.22) across all metrics, establishing a significant lead. The advantage is pronounced on complex multimodal landscapes (e.g., F5-D3), where GeM-EA reduces offline error by an order of magnitude compared to MLO. Figure 2. Online Error convergence trajectories on the last ten environments of SDDObenchF4-D4 Table 1. The Mean and Standard Deviation of EofflineE_offline on SDDObench Instance Drift DDEAs SDDEAs Proposed T-DDEA BDDEA-LDG MLO DETO DSE-MFS SAEF-1GP DASE GeM-EA F1 D1 6.82e+01± 2.15e-01 (+) 6.77e+01± 1.41e-02 (+) 6.82e+01± 1.50e-01 (+) 6.79e+01± 2.29e-01 (+) 6.77e+01± 2.37e-02 (+) 6.75e+01± 4.94e-03 (+) 6.76e+01± 1.90e-01 (+) 6.60e+01± 2.70e-01 D2 6.70e+01± 2.03e-01 (+) 6.78e+01± 6.72e-03 (+) 6.71e+01± 8.00e-02 (+) 6.66e+01± 4.08e-02 (+) 6.53e+01± 1.95e-02 (+) 6.76e+01± 4.61e-03 (+) 6.63e+01± 1.20e-01 (+) 6.45e+01± 2.30e-01 D3 6.59e+01± 1.97e-01 (+) 6.52e+01± 2.05e-02 (+) 6.60e+01± 1.20e-01 (+) 6.54e+01± 1.14e-01 (+) 6.56e+01± 9.93e-03 (+) 6.75e+01± 6.97e-03 (+) 6.48e+01± 2.10e-01 (+) 6.25e+01± 6.00e-01 D4 6.51e+01± 1.85e-01 (+) 6.46e+01± 2.20e-02 (+) 6.51e+01± 9.00e-02 (+) 6.45e+01± 9.70e-02 (+) 6.50e+01± 7.34e-03 (+) 6.56e+01± 2.33e-03 (+) 6.46e+01± 1.00e-01 (+) 6.31e+01± 1.10e-01 D5 6.53e+01± 1.92e-01 (+) 6.47e+01± 2.14e-02 (+) 6.53e+01± 1.10e-01 (+) 6.48e+01± 2.47e-01 (+) 6.51e+01± 2.61e-02 (+) 6.46e+01± 2.56e-03 (+) 6.48e+01± 1.00e-01 (+) 6.29e+01± 1.30e-01 F2 D1 6.82e+01± 2.11e-01 (+) 6.77e+01± 4.28e-03 (≈) 6.82e+01± 1.80e-01 (+) 6.79e+01± 2.41e-01 (+) 6.77e+01± 6.33e-02 (≈) 6.75e+01± 1.84e-03 (≈) 6.76e+01± 1.50e-01 (≈) 6.71e+01± 3.20e-01 D2 6.68e+01± 1.94e-01 (+) 6.76e+01± 6.27e-01 (+) 6.68e+01± 2.50e-01 (+) 6.76e+01± 3.05e-01 (+) 6.74e+01± 3.70e-01 (+) 6.76e+01± 4.09e-03 (+) 6.71e+01± 2.10e-01 (+) 6.51e+01± 3.80e-01 D3 6.66e+01± 1.88e-01 (+) 6.65e+01± 4.85e-01 (+) 6.71e+01± 3.20e-01 (+) 6.67e+01± 5.95e-02 (+) 6.63e+01± 3.12e-01 (+) 6.75e+01± 5.46e-03 (+) 6.61e+01± 7.90e-01 (+) 6.41e+01± 5.90e-01 D4 6.64e+01± 1.96e-01 (≈) 6.66e+01± 3.44e-01 (≈) 6.79e+01± 4.10e-01 (+) 6.63e+01± 1.59e-01 (≈) 6.66e+01± 1.04e+00 (≈) 6.76e+01± 3.53e-03 (+) 6.66e+01± 7.20e-01 (≈) 6.61e+01± 7.20e-01 D5 6.77e+01± 2.08e-01 (+) 6.63e+01± 2.26e-01 (≈) 6.63e+01± 2.80e-01 (≈) 6.61e+01± 4.19e-02 (≈) 6.69e+01± 9.16e-01 (+) 6.75e+01± 4.63e-03 (+) 6.70e+01± 8.40e-01 (+) 6.58e+01± 7.90e-01 F3 D1 6.82e+01± 2.15e-01 (+) 6.77e+01± 1.82e-02 (+) 6.82e+01± 1.60e-01 (+) 6.79e+01± 1.01e-01 (+) 6.78e+01± 1.83e-02 (+) 6.75e+01± 5.27e-03 (+) 6.76e+01± 2.10e-01 (+) 6.62e+01± 2.10e-01 D2 6.66e+01± 1.90e-01 (+) 6.73e+01± 1.33e-02 (+) 6.66e+01± 2.10e-01 (+) 6.61e+01± 1.83e-01 (+) 6.37e+01 ± 1.10e-02 (-) 6.76e+01± 3.03e-03 (+) 6.60e+01± 1.50e-01 (≈) 6.54e+01± 3.70e-01 D3 6.35e+01± 1.72e-01 (+) 6.33e+01± 1.03e-01 (+) 6.35e+01± 1.50e-01 (+) 6.29e+01± 3.90e-02 (+) 6.35e+01± 6.56e-02 (+) 6.35e+01 ± 3.99e-03 (+) 6.28e+01± 9.00e-02 (+) 6.18e+01± 2.70e-01 D4 6.33e+01± 1.65e-01 (+) 6.31e+01± 1.16e-01 (+) 6.32e+01± 1.80e-01 (+) 6.25e+01± 8.72e-02 (+) 6.34e+01 ± 3.74e-02 (+) 6.35e+01 ± 3.53e-03 (+) 6.28e+01± 1.20e-01 (+) 6.14e+01± 1.90e-01 D5 6.35e+01± 1.68e-01 (+) 6.30e+01± 9.22e-02 (+) 6.33e+01± 1.40e-01 (+) 6.28e+01± 2.15e-01 (+) 6.34e+01± 1.41e-01 (+) 6.35e+01± 6.15e-03 (+) 6.28e+01± 1.80e-01 (+) 6.12e+01± 1.60e-01 F4 D1 1.86e+02± 4.85e+00 (+) 4.21e+00± 3.41e-02 (+) 8.79e+01± 2.85e+00 (+) 1.08e+02± 3.34e+00 (+) 6.73e+01± 2.88e-01 (+) 5.04e+01± 2.80e-01 (+) 1.50e-01± 0.03e-01 (+) 6.00e-02± 1.00e-02 D2 1.50e+02± 3.76e+00 (+) 2.15e+01± 4.06e-01 (+) 1.01e+02± 3.40e+00 (+) 9.87e+01± 2.41e+00 (+) 4.35e+01± 3.09e-01 (+) 6.22e+01± 1.30e+00 (+) 1.46e+01± 3.38e+00 (+) 6.00e+00± 9.60e-01 D3 1.21e+02± 3.15e+00 (+) 2.20e+01± 2.47e-01 (+) 1.05e+02± 4.15e+00 (+) 8.62e+01± 2.74e+00 (+) 4.28e+01 ± 3.71e+00 (+) 6.08e+01± 2.90e-01 (+) 1.71e+01± 3.73e+00 (+) 5.35e+00± 5.80e-01 D4 1.13e+02± 2.84e+00 (+) 4.21e+01± 4.92e-01 (+) 1.19e+02± 5.20e+00 (+) 7.82e+01± 2.83e+00 (+) 5.45e+01± 3.39e+00 (+) 7.23e+01± 6.43e-01 (+) 2.99e+01± 5.19e+00 (+) 1.01e+01± 1.34e+00 D5 1.20e+02± 3.02e+00 (+) 4.01e+01± 1.09e+00 (+) 1.17e+02± 4.80e+00 (+) 8.40e+01± 1.69e+00 (+) 6.24e+01± 2.56e+00 (+) 7.40e+01± 1.67e-01 (+) 3.77e+01± 4.83e+00 (+) 1.37e+01± 2.24e+00 F5 D1 4.20e+03± 1.05e+02 (+) 1.31e+02± 3.15e+00 (+) 4.56e+03± 1.45e+02 (+) 6.34e+03± 1.78e+02 (+) 2.81e+03± 1.02e+02 (+) 2.09e+03± 1.20e+01 (+) 4.45e+01± 7.56e+00 (-) 5.62e+01± 6.30e+00 D2 4.57e+03± 1.14e+02 (+) 6.86e+02 ± 1.65e+01 (+) 4.84e+03± 1.68e+02 (+) 5.00e+03± 1.35e+02 (+) 2.01e+03± 7.83e+01 (+) 3.68e+03± 4.98e+01 (+) 5.80e+02± 1.50e+02 (+) 3.65e+02± 4.16e+01 D3 5.45e+03± 1.36e+02 (+) 6.39e+02 ± 1.76e+01 (+) 5.92e+03± 2.10e+02 (+) 4.73e+03± 1.11e+02 (+) 2.85e+03± 1.80e+02 (+) 2.95e+03± 2.51e+01 (+) 4.60e+02± 4.30e+01 (+) 2.59e+02± 3.05e+01 D4 5.36e+03± 1.34e+02 (+) 1.72e+03± 5.47e+01 (+) 7.13e+03± 2.55e+02 (+) 4.42e+03± 9.52e+01 (+) 2.99e+03± 9.44e+01 (+) 3.90e+03± 7.48e+01 (+) 1.44e+03± 3.41e+02 (+) 7.12e+02± 1.46e+02 D5 5.46e+03± 1.37e+02 (+) 1.77e+03± 5.76e+01 (+) 7.07e+03± 2.40e+02 (+) 4.62e+03± 1.37e+02 (+) 2.91e+03 ± 1.34e+02 (+) 4.04e+03± 1.74e+01 (+) 1.58e+03± 3.91e+02 (+) 8.18e+02± 1.35e+02 F6 D1 2.10e+01± 5.24e-01 (+) 1.15e+01± 3.27e-02 (+) 2.11e+01± 4.50e-01 (+) 2.07e+01± 4.83e-01 (+) 2.07e+01 ± 2.21e-02 (+) 2.03e+01± 6.80e-03 (+) 6.58e+00± 3.10e-01 (-) 7.67e+00± 3.10e-01 D2 2.11e+01± 5.28e-01 (+) 1.71e+01± 5.22e-02 (+) 2.12e+01± 5.10e-01 (+) 2.07e+01± 4.33e-01 (+) 2.03e+01± 1.01e-01 (+) 2.08e+01± 1.60e-02 (+) 1.24e+01± 3.80e-01 (+) 1.05e+01± 3.10e-01 D3 2.11e+01± 5.29e-01 (+) 1.85e+01± 4.45e-02 (+) 2.12e+01± 3.80e-01 (+) 2.06e+01± 4.33e-01 (+) 2.01e+01 ± 2.38e-02 (+) 2.07e+01± 1.11e-02 (+) 1.78e+01± 5.40e-01 (+) 1.71e+01± 5.20e-01 D4 2.13e+01± 5.32e-01 (+) 1.86e+01± 1.15e-02 (+) 2.13e+01± 4.20e-01 (+) 2.09e+01± 5.95e-01 (+) 2.07e+01± 2.81e-02 (+) 2.09e+01± 8.02e-03 (+) 1.80e+01± 4.60e-01 (+) 1.75e+01± 2.70e-01 D5 2.12e+01± 5.31e-01 (+) 1.85e+01± 3.01e-02 (+) 2.13e+01± 4.90e-01 (+) 2.08e+01± 5.42e-01 (+) 2.07e+01± 4.91e-02 (+) 2.09e+01± 7.28e-03 (+) 1.80e+01± 3.70e-01 (+) 1.73e+01± 2.20e-01 F7 D1 1.02e+00± 2.55e-02 (+) 5.28e-01± 7.81e-03 (-) 1.02e+00± 1.50e-02 (+) 1.00e+00± 1.81e-02 (+) 1.00e+00± 9.01e-04 (+) 1.01e+00± 9.64e-04 (+) 4.60e-01± 1.90e-01 (-) 5.90e-01± 9.00e-02 D2 1.02e+00± 2.56e-02 (+) 8.90e-01 ± 3.38e-03 (+) 1.02e+00± 2.10e-02 (+) 1.00e+00± 1.74e-02 (+) 1.01e+00 ± 1.88e-04 (+) 1.01e+00± 9.16e-04 (+) 6.50e-01± 7.00e-02 (-) 7.50e-01± 8.00e-02 D3 1.02e+00± 2.56e-02 (+) 9.99e-01± 4.54e-04 (+) 1.02e+00± 1.80e-02 (+) 9.96e-01± 1.53e-02 (+) 1.01e+00± 1.18e-03 (+) 1.01e+00± 1.08e-03 (+) 1.00e+00± 1.00e-02 (+) 9.20e-01± 4.00e-02 D4 1.03e+00± 2.57e-02 (+) 9.47e-01± 2.46e-03 (≈) 1.03e+00± 1.20e-02 (+) 1.00e+00± 1.98e-02 (+) 1.01e+00 ± 9.93e-04 (+) 1.01e+00± 1.33e-03 (+) 9.60e-01± 2.00e-02 (+) 9.40e-01± 2.00e-02 D5 1.03e+00± 2.57e-02 (+) 9.41e-01± 9.05e-03 (-) 1.03e+00± 1.50e-02 (+) 1.00e+00± 1.38e-02 (+) 1.01e+00 ± 1.59e-03 (+) 1.02e+00± 5.46e-04 (+) 9.60e-01± 2.00e-02 (-) 9.80e-01± 1.00e-02 F8 D1 1.66e+02± 4.14e+00 (+) 1.03e+02± 7.45e-01 (+) 1.85e+02± 5.45e+00 (+) 1.79e+02± 4.57e+00 (+) 1.70e+02± 9.17e-01 (+) 1.56e+02± 3.04e-01 (+) 7.33e+01± 7.72e+00 (-) 8.93e+01± 5.76e+00 D2 1.74e+02± 4.35e+00 (+) 1.25e+02± 7.33e-01 (+) 1.88e+02± 6.12e+00 (+) 1.79e+02± 5.07e+00 (+) 1.53e+02± 1.18e+00 (+) 1.69e+02 ± 2.76e-01 (+) 1.16e+02± 5.11e+00 (+) 1.10e+02± 1.70e+00 D3 1.72e+02± 4.31e+00 (+) 1.28e+02± 3.83e-02 (+) 1.86e+02± 5.85e+00 (+) 1.71e+02± 4.49e+00 (+) 1.57e+02 ± 6.61e-01 (+) 1.65e+02± 2.36e-01 (+) 1.24e+02± 3.43e+00 (+) 1.17e+02± 3.79e+00 D4 1.82e+02± 4.55e+00 (+) 1.43e+02± 1.01e+00 (+) 1.97e+02± 7.10e+00 (+) 1.70e+02± 4.88e+00 (+) 1.62e+02± 1.63e+00 (+) 1.75e+02± 6.12e-01 (+) 1.21e+02± 7.42e+00 (+) 1.15e+02± 2.06e+00 D5 1.83e+02± 4.58e+00 (+) 1.42e+02± 9.40e-01 (+) 1.99e+02± 6.90e+00 (+) 1.74e+02± 4.50e+00 (+) 1.67e+02± 1.68e+00 (+) 1.75e+02± 2.45e-01 (+) 1.36e+02± 6.03e+00 (+) 1.22e+02± 2.24e+00 +/≈/−+/≈/- 39/1/0 34/4/2 39/1/0 38/2/0 37/2/1 39/1/0 31/3/6 NA Average Rank 6.71 3.50 7.06 5.01 4.55 5.50 2.44 1.22 3.2.2. Dynamic Tracking Behavior Analysis. Figure 2 illustrates the online error trajectories. Following a drift, GeM-EA exhibits a “cliff-like” convergence profile, dropping almost vertically and stabilizing at a low error bound. This rapid recovery confirms the effectiveness of the bi-level meta-learned prior. Furthermore, thanks to the analytical solution for weights and linear residuals, GeM-EA maintains an O(N)O(N) computational complexity, running significantly faster than GP-based methods like SAEF-1GP and comparable to the fastest baselines. 3.3. Ablation & Sensitivity & Efficiency Due to strict space constraints, comprehensive ablation studies confirming the indispensability of each algorithmic component (BML, RAS, and GRE), detailed parameter sensitivity analyses on P and τ, and mathematical proofs of GeM-EA’s linear computational complexity O(N)O(N) alongside wall-clock runtime comparisons are presented in the Supplementary Appendix 2 and 3. 4. Conclusion This paper presents GeM-EA, a meta-learning enhanced evolutionary algorithm for Streaming Data-Driven Optimization (SDDO) under concept drift. By synergizing bi-level surrogate adaptation and decoupling structural and non-structural parameters to alleviate gradient conflicts, together with a multi-island generative replay mechanism, GeM-EA achieves rapid optima tracking while effectively mitigating negative transfer. SDDObench experiments confirm its superior convergence, cliff-like recovery, and linear complexity. Future work will leverage automated design principles and unified benchmarking (Qiu et al., 2026; Ma et al., 2025; Guo et al., 2025) to scale GeM-EA toward higher-dimensional, many-objective environments with fully autonomous algorithm design capabilities. Acknowledgements.This work was supported in part by the Guangdong Provincial Natural Science Foundation for Outstanding Youth Team Project (Grant No. 2024B1515040010), in part by Guangzhou Science and Technology Elite Talent Leading Program for Basic and Applied Basic Research (Grant No. SL2024A04J01361), in part by the Fundamental Research Funds for the Central Universities (Grant No. 2025ZYGXZR027). References X. Chen, J. Wang, and K. Xie (2021) TrafficStream: a streaming traffic flow forecasting framework based on graph neural networks and continual learning. IJCAI. Cited by: §1. C. Finn, P. Abbeel, and S. Levine (2017) Model-agnostic meta-learning for fast adaptation of deep networks. In International conference on machine learning, p. 1126–1135. Cited by: §1. C. Finn, A. Rajeswaran, S. Kakade, and S. Levine (2019) Online meta-learning. In International conference on machine learning, p. 1920–1930. Cited by: §1. Y. Gong, Y. Zhong, and H. Huang (2023) Offline data-driven optimization at scale: a cooperative coevolutionary approach. IEEE Transactions on Evolutionary Computation 28 (6), p. 1809–1823. Cited by: §1. H. Guo, Z. Ma, Y. Ma, X. Zhang, W. Chen, and Y. Gong (2025) DesignX: human-competitive algorithm designer for black-box optimization. arXiv preprint arXiv:2505.17866. Cited by: §4. P. Huang, H. Wang, and Y. Jin (2021) Offline data-driven evolutionary optimization based on tri-training. Swarm and evolutionary computation 60, p. 100800. Cited by: §3.1. J. Li, Z. Zhan, C. Wang, H. Jin, and J. Zhang (2020) Boosting data-driven evolutionary algorithm with localized data generation. IEEE Transactions on Evolutionary Computation 24 (5), p. 923–937. Cited by: §3.1. K. Li, R. Chen, and X. Yao (2023) A data-driven evolutionary transfer optimization for expensive problems in dynamic environments. IEEE Transactions on Evolutionary Computation. Cited by: §3.1. W. Luo, R. Yi, B. Yang, and P. Xu (2018) Surrogate-assisted evolutionary framework for data-driven dynamic optimization. IEEE Transactions on Emerging Topics in Computational Intelligence 3 (2), p. 137–150. Cited by: §3.1. Z. Ma, Y. Gong, H. Guo, W. Qiu, S. Ma, H. Lian, J. Zhan, K. Chen, C. Wang, Z. Huang, Z. Huang, G. Peng, R. Cheng, and Y. Ma (2025) MetaBox-v2: a unified benchmark platform for meta-black-box optimization. In Advances in Neural Information Processing Systems, Cited by: §4. E. Osekowska, H. Johnson, and B. Carlsson (2017) Maritime vessel traffic modeling in the context of concept drift. Transportation research procedia 25, p. 1457–1476. Cited by: §1. J. Park and I. W. Sandberg (1991) Universal approximation using radial-basis-function networks. Neural computation 3 (2), p. 246–257. Cited by: §2. M. L. Peixoto, E. Mota, A. H. Maia, W. Lobato, M. A. Salahuddin, R. Boutaba, and L. A. Villas (2023) FogJam: a fog service for detecting traffic congestion in a continuous data stream vanet. Ad Hoc Networks 140, p. 103046. Cited by: §1. W. Qiu, H. Guo, Z. Ma, J. Zhang, and Y. Gong (2026) Automated algorithm design for black-box optimization: progress and challenges. Chinese Journal of Computers 49 (4). External Links: Document Cited by: §4. S. Ravi and H. Larochelle (2017) Optimization as a model for few-shot learning. In International conference on learning representations, Cited by: §1. O. Vinyals, C. Blundell, T. Lillicrap, D. Wierstra, et al. (2016) Matching networks for one shot learning. Advances in neural information processing systems 29. Cited by: §1. G. Widmer and M. Kubat (1996) Learning in the presence of concept drift and hidden contexts. Machine learning 23 (1), p. 69–101. Cited by: §1. C. Yang, J. Ding, Y. Jin, and T. Chai (2023) A data stream ensemble assisted multifactorial evolutionary algorithm for offline data-driven dynamic optimization. Evolutionary Computation 31 (4), p. 433–458. Cited by: §3.1. H. Zhang, J. Ding, L. Feng, K. C. Tan, and K. Li (2024) Solving expensive optimization problems in dynamic environments with meta-learning. IEEE transactions on cybernetics. Cited by: §1, §3.1. Y. Zhong and Y. Gong (2025) Data-driven evolutionary computation under continuously streaming environments: a drift-aware approach. IEEE Transactions on Evolutionary Computation. Cited by: §1, §3.1. Y. Zhong, T. Huang, X. Xiao, and Y. Gong (2026) TRACE: a generalizable drift detector for streaming data-driven optimization. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 40, p. 28910–28918. Cited by: 1st item. Y. Zhong, X. Wang, Y. Sun, and Y. Gong (2024) Sddobench: a benchmark for streaming data-driven optimization with concept drift. In Proceedings of the Genetic and Evolutionary Computation Conference, p. 59–67. Cited by: §1, §3.1.