Learning to Evolve for Optimization via Stability-Inducing Neural Unrolling

A research team led by Jiaxin Gao has introduced 'Learning to Evolve' (L2E), a novel framework designed to overcome a fundamental challenge in AI-driven optimization. Traditional evolutionary algorithms rely on manually engineered heuristics, limiting their adaptability. While learned optimizers offer more flexibility, their unconstrained update rules often lead to unstable dynamics and poor generalization. L2E addresses this 'plasticity-stability dilemma' through a bilevel meta-optimization approach that learns evolutionary search via stability-inducing neural unrolling, promising more robust and adaptable optimization tools.

The technical core of L2E reformulates population evolution as an unrolled fixed-point iteration using a structured neural operator. An inner loop imposes a stability-biased update structure, while an outer loop meta-trains the operator to produce effective search trajectories across various tasks. To balance exploration and refinement, the framework employs a gradient-derived composite solver that adaptively fuses learned evolutionary proposals with proxy numerical guidance in a differentiable manner. Extensive testing on synthetic benchmarks and real-world control tasks demonstrates that L2E achieves significant optimization performance, scales effectively to high-dimensional problems, and shows strong zero-shot transfer capabilities across diverse test distributions, marking a step toward more general and reliable AI optimization systems.