MapElitesAutoMLArchiveEntry

Models & Types · AiDotNet.AutoML

Describes one immutable elite model specification retained by MAP-Elites AutoML.

For Beginners

After an AutoML run that uses the MAP-Elites strategy, the Archive property of the AutoML model lists one entry like this per occupied cell. Each entry says which model type won that cell, the exact hyperparameters it used, the validation score it earned, and where it sits in the descriptor grid. For example, one entry might read: model type RandomForestRegression, parameters NumberOfTrees = 200 and MaxDepth = 6, score 0.91, descriptors model_family = 2 and configuration_complexity = 0.55, cell bins [2, 2]. Browse the archive when you want to understand trade-offs (a simpler model that scores nearly as well) or to retrain a specific elite yourself rather than only the single best model that AutoML returns.

How It Works

MAP-Elites AutoML keeps the best validated model specification in each behavior cell, where the cell is defined by the model family and the normalized configuration complexity (see MapElitesAutoMLOptions). This type is the public, read-only projection of one such cell: it exposes the specification and its scores, never a trained model instance, so inspecting a large archive does not keep one live model per cell. All collections are defensive copies; mutating the source dictionaries after construction has no effect on the entry.

Properties

Property Summary
IReadOnlyDictionary<string, double> Descriptors { get; } Gets the behavior descriptors computed without final-test data.
IReadOnlyDictionary<string, object> Parameters { get; } Gets a defensive read-only copy of the elite hyperparameters.
IReadOnlyList<int> CellBins { get; } Gets one zero-based bin index per descriptor.
Type ModelType { get; } Gets the open generic or concrete model type represented by this elite.
double Score { get; } Gets the validation score used as archive quality.
long EvaluationId { get; } Gets the stable engine evaluation identifier.
string SpecificationId { get; } Gets the canonical identity of the model specification.