Notes
🔍

FLAML

Most AutoML libraries assume you can spare a few hours of compute. FLAML assumes you cannot. It is Microsoft Research's answer to the question of what an automated search should do when the honest budget is one minute on a laptop CPU rather than an afternoon on a cluster.

Cost-frugal search

The name stands for Fast and Lightweight AutoML, and the economy comes from treating training cost as part of the search problem rather than as an afterthought. Two algorithms carry it:

  • Cost-Frugal Optimisation (CFO). A randomised local search that starts from the cheapest configuration — shallow trees, few iterations, small samples — and moves outward, entering expensive regions only when the cheap ones stop paying.
  • BlendSearch. CFO blended with a global Bayesian optimiser, which keeps the low-cost bias while still escaping a poor local basin.

Using it

The interface is a task and a time budget:

from flaml import AutoML
automl = AutoML()
automl.fit(X_train, y_train, task='classification', time_budget=60)

Where it fits

The estimator pool leads with gradient boosting — LightGBM and XGBoost — which is why FLAML is strongest on tabular problems. Beyond the wrapper, flaml.tune exposes the same optimisers as a standalone hyperparameter tuner for anything you can score, including model fine-tuning, and it runs distributed on Ray. Microsoft uses it inside the tuning layer of Azure AutoML, and it is the reference choice whenever the time limit, not the search space, is the binding constraint.

FLAML (Fast and Lightweight AutoML) is an open-source Python library from Microsoft Research that finds accurate models under tight time and compute budgets. Its cost-frugal optimisation and BlendSearch algorithms begin at cheap configurations and spend only where accuracy improves, instead of sampling the search space blindly. It matters because it makes AutoML practical in minutes on a laptop, and its tuner, flaml.tune, works independently of the AutoML wrapper.

Imagine shopping for a car on a strict budget. You do not test-drive the Ferraris first. You start with what you can afford, and you only look at pricier options if the cheap ones clearly fall short.

FLAML is a free tool, made by Microsoft, that builds prediction models that way. Other tools try lots of heavy, slow options in search of a good model, which can take hours on an expensive machine. FLAML begins with the quick, cheap options and spends more computing time only when the extra cost actually buys better accuracy. You hand it your data and tell it how many minutes it is allowed, and it works inside that limit — which is why it runs happily on an ordinary laptop.