DNA-Style Hyperparameter Evolution Results

This devlog covers the design and initial outcomes of the genetics-inspired hyperparameter evolution work in AgentFarm.

Related tracking issue: Issue #22

What Changed

The core design shift was to model tunable parameters as explicit genes inside a typed chromosome instead of using free-floating config values.

• Added HyperparameterChromosome and HyperparameterGene with explicit bounds and validation.
• Implemented encode/decode support for genes, including quantized and log-scale handling for learning-rate style ranges.
• Added genetic operators for selection, crossover, and mutation.
• Added a generation-based evolution runner that persists lineage and summaries for reproducibility.

Design details:

• Hyperparameter Chromosome Design
• Convergence Experiment Notes

Why This Design

Using an explicit chromosome structure gives us:

• a clear contract for what can evolve
• deterministic, auditable experiment artifacts
• safer mutation/crossover behavior with bounded genes
• easier extension to additional parameters in later iterations

This keeps the evolutionary logic intentional instead of ad hoc.

What the Initial Runs Showed

Two comparison runs were executed using different selection and mutation settings:

1. Tournament + lower mutation pressure
2. Roulette + higher mutation pressure

High-level outcome:

• Tournament with lower mutation improved best fitness, but pushed the winning learning rate to the lower bound.
• Roulette with higher mutation preserved broader exploration but showed flat best fitness.

This is the expected exploration/exploitation trade-off showing up clearly in the data.

Overall Interpretation

The pipeline now works end-to-end and responds to different evolutionary settings in meaningful ways.

Current behavior suggests:

• lower mutation + stronger selection can find useful signal, but risks boundary collapse
• higher mutation + softer selection can keep diversity, but may wash out optimization progress

So the next practical tuning direction is:

• keep tournament-style selection pressure
• lower mutation pressure further
• add a soft lower-bound guard/penalty so improvements do not depend on boundary collapse

Artifacts

The current comparison outputs live under:

• experiments/evolution_convergence/run_tournament_mut020_g6
• experiments/evolution_convergence/run_roulette_mut040_g6

Each run includes:

• evolution_generation_summaries.json
• evolution_lineage.json
• run_manifest.json
• hyperparameter_evolution.png

Closing Note

This is a strong foundation milestone: the representation, operators, and experiment loop are in place and reproducible. The next stage is focused tuning and adding more evolvable genes with the same discipline.