limen.experiment.param_search

Canonical docs

Advanced Search

In the research projects

Both projects run search over recorded observations. The Python project constructs strategies over ParamDomain; Compiled project resolves them from its declaration. The assertions compare completed candidate identities and keep candidate-generation seeds separate from model randomness.

What this package owns

Owns the pluggable search strategies the advanced experiment path uses to enumerate parameter permutations, plus the registry that names them. Does not own the parameter domain (limen.experiment.param_domain), the search queue (limen.experiment.msq), or the loop that drives them.

Key entry points

Entry point	Use case	Notes
`SearchStrategy`	Abstract base for strategies	Iterators that yield `dict[str, Any]` combinations and observe domain mutations
`GridStrategy`	Exhaustive grid over the domain	Deterministic full sweep
`RandomStrategy`	Sampled combinations	Accepts a `seed` for reproducibility
`SobolStrategy`	Scrambled Sobol candidate sampling	Requires a nonnegative integer `seed`; discrete projection is not exhaustive
`STRATEGY_REGISTRY`	Name-to-class lookup	Resolves the strategy named in a manifest

Adjacent modules

limen.experiment.param_domain.ParamDomain supplies the domain strategies iterate; strategies register as observers so they react to domain mutations.
limen.experiment (UEL, MSQ, reducers) drives the strategy and feeds feedback back into the domain.

Quick orientation

Module or family	Responsibility
`search_strategy.py`	SearchStrategy abstract base
`grid_strategy.py`	Exhaustive grid
`random_strategy.py`	Random sampling
`sobol_strategy.py`	Seeded scrambled Sobol sampling
`registry.py`	STRATEGY_REGISTRY

text · README.md

1param_search/2├── README.md3├── __init__.py4├── _grid_permutation.py5├── grid_strategy.py6├── random_strategy.py7├── registry.py8├── search_strategy.py9└── sobol_strategy.py

Merged source PR · f78356f4Package layout · Limen 6.0.1 · current package validation · package source

Things to know

A strategy is an iterator that yields one parameter dict per proposal. Grid stops at exhaustion; sampled strategies can fail when bounded duplicate retries find no novel combination.
Strategies hold a reference to their ParamDomain and register as observers, so reducers that mutate the domain mid-run change what the strategy yields next.
RandomStrategy is only reproducible when constructed with an explicit seed.
SobolStrategy(domain, seed=42) is also exported from this package; its mapping, limits and checkpoint contract are documented in Advanced Search.