MapElitesAutoMLOptions
Options & Configuration · AiDotNet.Configuration
Configures quality-diversity search when AutoML uses the MAP-Elites strategy.
For Beginners
Ordinary AutoML keeps chasing the single best score, which tends to funnel
every trial into one model family and one region of settings. MAP-Elites instead lays out a grid whose
axes are the model family and how complex a configuration is, and it keeps the best-scoring specification
found in each grid cell, so at the end you can see the strongest simple linear model, the strongest
mid-complexity tree ensemble, and so on, while the overall winner is still returned normally. These
settings control that exploration: how many random seed configurations to try first, how finely to slice
complexity, how often to mutate or explore instead of refining known elites, and how many elites may
inspire one new proposal. Attach an instance through
MapElites when SearchStrategy is
AutoMLSearchStrategy.MapElites. The defaults suit most tabular problems, so start by changing only
Seed for reproducibility and ComplexityBinCount for a coarser or finer archive.
How It Works
MAP-Elites retains the best validated model specification in each behavior cell instead of retaining only one globally best configuration. The archive stores specifications and scores, not trained model instances, so archive growth does not retain one live model per cell.
The algorithm follows Mouret and Clune, “Illuminating search spaces by mapping elites” (2015). Proposal
randomness is derived from Seed through stable per-candidate streams rather than a shared
global generator, so the search is reproducible whenever the underlying model training is, and duplicate
specifications are rejected before training so they never consume the trial budget.
Constructors
| Constructor | Summary |
|---|---|
MapElitesAutoMLOptions() |
Properties
| Property | Summary |
|---|---|
double ExplorationProbability { get; set; } |
Gets or sets the probability of proposing a fresh model family and configuration. |
double MutationProbability { get; set; } |
Gets or sets the probability that a mutable parameter is resampled. |
int ArchiveCapacity { get; set; } |
Gets or sets the maximum occupied archive cells, or zero to use the complete descriptor grid. |
int ComplexityBinCount { get; set; } |
Gets or sets the number of bins for normalized configuration complexity. |
int InitialPopulationSize { get; set; } |
Gets or sets the minimum number of random seed configurations. |
int InspirationCount { get; set; } |
Gets or sets how many elite configurations may inform one proposal. |
int IslandCount { get; set; } |
Gets or sets the number of independent island archives; one, the default, keeps a single population. |
int MaxProposalMultiplier { get; set; } |
Gets or sets the proposal allowance per expensive trial attempt. |
int MigrantsPerIsland { get; set; } |
Gets or sets the maximum elites copied out of each source island in one migration round. |
int MigrationInterval { get; set; } |
Gets or sets the committed batches between island migrations; zero, the default, disables migration. |
ulong Seed { get; set; } |
Gets or sets the stable root seed used to derive proposal-local random streams. |