New model learns to optimize unseen problems without retraining

Optimization—the process of finding the best solution to a problem—is fundamental across machine learning and engineering. Yet current optimizers often struggle when faced with new problem types they haven't seen before, typically requiring extensive tuning to work effectively.

A team of researchers has introduced a Pretrained Optimization Model (POM) designed to generalize across unfamiliar optimization challenges, potentially reducing the need for manual adjustment and adaptation.

How it works#

Rather than building an optimizer from scratch for each new task, POM learns patterns from training across many different optimization problems. This broad exposure gives it the ability to apply its learned knowledge directly to novel tasks—a capability researchers call zero-shot optimization. When more performance is needed, the model can be refined using just a handful of sample evaluations, a process known as few-shot fine-tuning.

Empirical results#

Testing on the Black-Box Optimization Benchmark (BBOB) and real-world robot control tasks, POM outperformed leading black-box optimization methods, with particularly strong gains on high-dimensional problems. The model also proved robust across variations in task type, problem dimension, population size, and optimization budget.

Why it matters#

Traditional black-box optimizers require careful hyperparameter tuning every time they encounter a new problem. By learning generalizable optimization strategies across diverse tasks, POM sidesteps this bottleneck. Even modest fine-tuning with limited samples delivered notable improvements, suggesting the approach could accelerate optimization in fields ranging from engineering design to machine learning.

The work is described in a preprint submitted to arXiv in May 2024, with code made available to the community.