ProgramMetricAggregator

Models & Types · AiDotNet.Evolution.Programs.Metrics

Turns a dictionary of evaluator metrics into the single scalar quality an archive ranks on.

For Beginners

Your evaluation code measures several things about a candidate program, but the search can only rank on one number. This class does that conversion and, just as importantly, tells you what it did. Construct one with default options and call Aggregate; read Value for the score, check HasFiniteValue before reporting it as a completed evaluation, and read Issues when the score is not what you expected.

How It Works

The default strategy reproduces the reference OpenEvolve rule (openevolve/utils/metrics_utils.py``get_fitness_score, lines 72-118) exactly: an empty dictionary scores zero; a metric literally named combined_score is converted and returned outright; otherwise the numeric metrics that are not archive feature dimensions are averaged; and when excluding the feature dimensions leaves nothing, every numeric metric is averaged instead. Booleans never count as scores - Python treats True as 1, so averaging a timeout flag in would give a crashed program a mid-range fitness - and NaN values are dropped, while infinities are not.

Two things are deliberately different. First, every value the rule discards is returned as a ProgramMetricIssue instead of vanishing, so an evaluator that reports its accuracy as the string "0.9" is distinguishable from one whose program genuinely scored zero. Second, text is converted only when it parses to a finite invariant-culture number, where Python’s float() would also accept the literals nan and inf; that keeps the outcome identical on every target framework. Beyond the reference rule, Weighted and Tchebycheff offer aggregations whose parameters are validated before a run begins rather than assumed.

Constructors

Constructor Summary
ProgramMetricAggregator(ProgramMetricAggregationOptions options) Initializes an aggregator.

Properties

Property Summary
EvolutionOptimizationDirection PreferredDirection { get; } Gets whether larger or smaller aggregated values are better under the configured strategy.

Methods

Method Summary
ProgramMetricAggregationOptions GetOptions() Gets an independent copy of the settings this aggregator was validated with.
ProgramMetricAggregationResult Aggregate(IReadOnlyDictionary<string, ProgramMetricValue> metrics) Collapses an all-numeric metrics dictionary into one scalar quality.
ProgramMetricAggregationResult Aggregate(IReadOnlyDictionary<string, double> metrics) Collapses an all-numeric metrics dictionary into one scalar quality.
ProgramMetricAggregationResult UpstreamFitnessScore(IReadOnlyDictionary<string, ProgramMetricValue> metrics, IEnumerable<string> featureDimensions) Applies the reference OpenEvolve fitness rule with no other configuration.

Fields

Field Summary
string AggregateMetricName The metric name reported on an issue that concerns the aggregation itself rather than one metric.