GeM-EA: A Generative and Meta-learning Enhanced Evolutionary Algorithm for Streaming Data-Driven Optimization
New meta-learning algorithm handles sudden environmental changes in real-time data streams, outperforming state-of-the-art methods.
A research team led by Yue Wu and Yue-Jiao Gong has introduced GeM-EA, a novel algorithm designed to tackle the complex challenge of Streaming Data-Driven Optimization (SDDO). In SDDO problems, data arrives continuously and the underlying optimization environment evolves over time, a phenomenon known as concept drift. This drift creates non-stationary landscapes that can render traditional optimization models obsolete. GeM-EA's core innovation is its unified approach, combining meta-learned surrogate adaptation with generative replay to maintain an effective evolutionary search process in these shifting conditions.
Upon detecting a concept drift, GeM-EA employs a bi-level meta-learning strategy. This allows the algorithm to rapidly initialize a new predictive model (surrogate) using knowledge priors relevant to the new environment, while a linear residual component captures broader global trends. Simultaneously, a multi-island evolutionary strategy leverages historical knowledge through a technique called generative replay, which helps accelerate the search for optimal solutions by preventing catastrophic forgetting of past useful strategies. Experimental results on established SDDO benchmarks demonstrate that GeM-EA achieves significantly faster adaptation and greater robustness compared to current state-of-the-art methods, effectively mitigating the issue of negative transfer that plagues simpler approaches during sudden environmental shifts.
The paper, accepted for presentation at the prestigious GECCO 2026 conference, addresses a critical gap in real-time optimization systems. Existing methods often rely on simplistic combinations of models or directly injecting old solutions, which can fail dramatically when the environment changes abruptly. By formally unifying meta-learning for quick model adjustment and generative techniques for preserving useful historical knowledge, GeM-EA provides a more resilient and efficient framework for dynamic optimization tasks found in areas like financial trading, supply chain logistics, and real-time system control.
- Uses bi-level meta-learning to rapidly initialize new surrogate models upon detecting concept drift, leveraging environment-relevant priors.
- Integrates generative replay within a multi-island evolutionary strategy to preserve and utilize historical knowledge, accelerating optimization.
- Demonstrated on benchmarks to achieve faster adaptation and improved robustness compared to state-of-the-art SDDO methods, mitigating negative transfer.
Why It Matters
Enables more reliable real-time AI for finance, logistics, and control systems that must adapt instantly to changing data streams.