godon-breeders

Autonomous breeder agents for optimization using metaheuristic search.

Architecture

Breeders are self-driving optimization agents that use Optuna ask/tell pattern for parameter search - further metaheuristics frameworks may follow. Effectuation and reconnaissance are executed as Windmill scripts on target systems.

The system follows an engine + strains architecture: the engine provides the generic optimization loop (algorithm diversity, guardrails, rollback, cooperation, metrics), while strains encapsulate domain-specific knowledge (parameter suggestion, validation).

Engine (engine/)

• BreederWorker: Generic optimization agent with lifecycle management, algorithm diversity across parallel workers, guardrail checking, and rollback support
• Communication: Cooperative trial sharing between breeders via Optuna database (probabilistic, best, worst, extremes strategies)
• BreederMetricsClient: Prometheus metrics pushing via Push Gateway
• Strain Loader: Dynamic loading and contract validation of strain modules

Strains (strains/)

Each strain provides domain-specific logic as a pluggable module:

• suggest_params(trial, settings) — parameter suggestion for Optuna trials
• validate_config(config) — configuration validation (preflight checks)

Effectuation (effectuation/)

Scripts that apply parameter changes to target systems. Each script follows the (context, targets, settings) interface contract:

• context — static breeder run configuration (credentials, URLs, playbook paths)
• targets — list of target systems to apply changes to
• settings — the optimizer's parameter suggestions for this trial

Available effectuators:

• SSH — applies configuration via Ansible playbooks over SSH
• HTTP — applies configuration via HTTP API calls (draft)

Reconnaissance (reconnaissance/)

Scripts that gather metrics to evaluate trial outcomes. Same (context, targets, settings) interface contract. Supports:

• Prometheus — multi-sample collection, stabilization waits, and aggregation
• HTTP — metric collection via HTTP GET with configurable stabilization, multi-sampling, and aggregation

Available Strains

linux_performance (strains/linux_performance/)

Optimizes Linux system parameters (sysctl, sysfs, cpufreq, ethtool) for improved performance. Supports network, memory, CPU, and custom optimization objectives via Prometheus metrics.

bench_greenhouse (strains/bench_greenhouse/)

Optimizes greenhouse climate simulation parameters. Supports multi-zone heating, ventilation, shading, CO2 injection, lighting, and irrigation. Designed for the godon-bench-greenhouse simulation container with HTTP-based effectuation and reconnaissance.

License

AGPL-3.0