ASHAOptimizer<T, TInput, TOutput>

Models & Types · AiDotNet.HyperparameterOptimization

Implements ASHA (Asynchronous Successive Halving Algorithm) for hyperparameter optimization.

Inheritance: HyperparameterOptimizerBase<T, TInput, TOutput> → ASHAOptimizer<T, TInput, TOutput>

How It Works

For Beginners: ASHA is an improved version of Hyperband that:

  • Doesn’t wait for all configurations at a level before promoting some
  • Can promote promising configurations as soon as they outperform enough peers
  • Is naturally suited for parallel/distributed training
  • Typically converges faster than synchronous Hyperband

Key concepts:

  • Rungs: Resource levels at which we evaluate (e.g., epochs 1, 3, 9, 27)
  • Promotion: Moving a configuration to train with more resources
  • Early Stopping: Killing configurations that aren’t competitive

ASHA uses the same exponential resource increase as Hyperband but allows asynchronous promotion based on relative performance.

Constructors

Constructor Summary
ASHAOptimizer(bool maximize, int maxResource, int reductionFactor, int minResource, double promotionThreshold, int? seed) Initializes a new instance of the ASHAOptimizer class.

Properties

Property Summary
List<int> Rungs { get; } Gets the resource levels (rungs) in this ASHA configuration.

Methods

Method Summary
Dictionary<int, RungStatistics> GetRungStatistics() Gets statistics about configurations at each rung.
Dictionary<string, object> SuggestNext(HyperparameterTrial<T> trial) Suggests the next hyperparameter configuration to try.
HyperparameterOptimizationResult<T> Optimize(Func<Dictionary<string, object>, T> objectiveFunction, HyperparameterSearchSpace searchSpace, int nTrials) Searches for the best hyperparameter configuration using ASHA.
ValueTuple<Dictionary<string, object>, T> GetBestConfiguration() Gets the best configuration found at the highest rung.