AutoML limits and specifications

Supported regions

AutoML is available in the following regions: China (Hangzhou), China (Shanghai), China (Beijing), China (Shenzhen), and China (Hong Kong).

Supported search algorithms

AutoML hyperparameter search supports the following algorithms and their typical scenarios.

• TPE (Tree-structured Parzen Estimator)

  TPE requires no additional dependencies and supports all search space types. It is the default HPO algorithm. Use it for complex, nonlinear, high-dimensional problems with high computational cost. It cannot discover relationships between parameters.

  Reference: Algorithms for Hyper-Parameter Optimization.

• GridSearch

  GridSearch evenly divides the search space into a grid and evaluates every combination to find the best one. Use it when the search space is small.

• Random

  Random generates hyperparameter combinations. Like grid search, it divides the search space into a grid, but each trial randomly selects one hyperparameter set. Use it for nonlinear, high-dimensional problems with high computational cost.

• Evolution

  Evolution initializes from the search space. In each generation, it selects the better-performing parameters and mutates them to produce the next set of candidate hyperparameters. It requires multiple trials. The logic is simple and easy to extend.

  The idea comes from Large-Scale Evolution of Image Classifiers.

• GP (Gaussian process)

  GP is a Bayesian optimization method that uses a Gaussian process to compute loss. As more data accumulates, the posterior distribution becomes more accurate and the optimization improves.

• PBT (Population Based Training)

  PBT is an asynchronous optimization algorithm for fixed compute resources. It optimizes a fixed number of models and their hyperparameters to improve performance, iterating on a single hyperparameter set to reach the optimum.

  The algorithm comes from Population Based Training of Neural Networks.

References

Create an experiment