odatse.algorithm.global_search module

class odatse.algorithm.global_search.Algorithm(info: Info, runner: Runner = None, domain=None, run_mode: str = 'initial')

Bases: AlgorithmBase

Algorithm class for global optimization using scipy.optimize routines.

The optimization method is selected by the method parameter in the [algorithm.global_search] section. Currently implemented:

• “DE” / “differential_evolution”: scipy.optimize.differential_evolution

• “shgo”: scipy.optimize.shgo

• “direct”: scipy.optimize.direct

• “dual_annealing”: scipy.optimize.dual_annealing

The per-method differences (aliases, seed and workers handling, callback signature, defaults, output files) are described by the module-level _METHODS table.

All other entries of the section are passed verbatim as arguments of the selected scipy routine; argument names the routine does not accept abort before the optimization starts.

MPI parallelization uses a master-worker layout over the algorithm communicator: algorithm rank 0 drives the scipy optimizer, whose workers hook scatters candidate points (for DE, a whole generation at a time) to all algorithm ranks; the other ranks run an evaluation-server loop, evaluating their share of the points with their own solver group. This composes with solver-side parallelism (nsolve): the total parallelism is algsize (points) x nsolve (per point). The direct and dual_annealing methods do not support parallel evaluation and run entirely on rank 0.

Initialize the Algorithm class.

Parameters:

• info (Info) – Information object containing algorithm settings.

• runner (Runner) – Runner object for submitting jobs.

• domain – Domain object defining the search space.

• run_mode (str) – Mode of running the algorithm.

__init__(info: Info, runner: Runner = None, domain=None, run_mode: str = 'initial') → None

Initialize the Algorithm class.

Parameters:

• info (Info) – Information object containing algorithm settings.

• runner (Runner) – Runner object for submitting jobs.

• domain – Domain object defining the search space.

• run_mode (str) – Mode of running the algorithm.

_initialize() → None

Set up initial state for a fresh run.

The global search does not use checkpointing, so this is a no-op.

_output_results()

Output the results of the optimization to files.

Every algorithm rank writes the history of its own function evaluations; the iteration history and the result summary exist only on rank 0, which drove the optimizer.

_post()

Post-process the results after optimization.

_prepare() → None

Algorithm-specific preparation, called after dispatch.

Override in subclasses to perform setup that must happen after the checkpoint state is established (e.g. initialising timer entries).

_run() → None

Run the global optimization.

Algorithm rank 0 drives the scipy optimizer; the other algorithm ranks serve function evaluations until rank 0 signals completion.

class odatse.algorithm.global_search._Method(func: Callable | None, requires: str | None, aliases: tuple, uses_seed: bool, supports_workers: bool, callback_style: str, defaults: dict, iter_file: str, iter_header: str)

Bases: object

Declarative description of one scipy.optimize global routine.

All per-method differences of the algorithm live here, so that adding a method amounts to adding one entry to the _METHODS table (plus tests and documentation); __init__, _run and _output_results are table-driven.