SMAC
SMAC is a hyperparameter and algorithm-configuration tool designed for the messy, expensive optimization problems that trip up simpler methods. Its name stands for Sequential Model-based Algorithm Configuration, and it has a long track record as one of the engines behind automated machine learning.
How it works
SMAC is a form of Bayesian optimization, but instead of the Gaussian process model most such tools use, it fits a random forest as its surrogate model. That choice is deliberate: random forests cope gracefully with search spaces that mix continuous, integer, categorical, and conditional parameters — settings that only matter when another setting takes a particular value. SMAC evaluates configurations sequentially, updates its model after each result, and uses it to propose the next configuration most likely to improve on the best found so far.
Why it matters in AutoML
SMAC targets exactly the hard cases automated systems face:
- It handles complex, structured configuration spaces that break Gaussian-process methods.
- It supports algorithm configuration across many problem instances, not just tuning on one dataset.
- It powers Auto-sklearn, using its budget wisely when each model evaluation is costly.
Maintained by the AutoML group at the University of Freiburg and available as SMAC3 in Python, it is the tool to reach for when your search space is large, mixed, and expensive to evaluate.
SMAC (Sequential Model-based Algorithm Configuration) is a Bayesian-optimization tool for hyperparameter tuning and algorithm configuration. It uses a random forest surrogate model, which handles the mixed continuous, categorical, and conditional search spaces that trip up Gaussian-process methods. Built for expensive, structured problems and configuration across many instances, it powers Auto-sklearn and is available as SMAC3 in Python.
Some machine-learning setups have a tangle of settings, where certain options only matter when others are switched on. Finding good combinations by trial and error is slow and costly. SMAC is a tool that tackles this smartly. It keeps a running model of which combinations tend to work, using it to make each new attempt more informed than the last, so it wastes little effort. Its particular strength is coping with tangled, mixed sets of options that simpler methods stumble on. Because it makes every expensive test count, SMAC has long been a trusted engine inside automated machine-learning systems.