API reference#

These pages are generated from the current source, including constructor signatures, methods, and existing docstrings. Generated API documentation is shared in English by both language guides. All 57 public algorithm constructors include parameter descriptions, types, defaults, and implementation notes. Prefer public module exports for imports; internal helpers are not stable interfaces.

Selection operator overview#

Import common operators from evomo.operators.selection. Objectives use minimization.

Interface	Input	Output and purpose
`non_dominate_rank(x, cv=None)`	`(N, M)` objectives and optional violations	Complete zero-based ranks of shape `(N,)`
`crowding_distance(costs, mask)`	Objectives and a boolean `(N,)` mask, or `None`	Crowding distances of shape `(N,)`
`nd_environmental_selection(x, f, topk, cv=None)`	Population, objectives, survivor count, and optional violations	`(selected_pop, selected_fit, selected_rank, selected_distance, selected_cv)`
`ref_vec_guided(x, f, v, theta)`	Population, objectives, reference vectors, and angle penalty parameter	`(next_pop, next_fit)`; unassociated slots may contain NaN
`get_non_dominate_backend(backend="torch")`	`"torch"` or `"triton"`	Backend module; does not change algorithms automatically

For unconstrained environmental selection, the fifth return value is None. topk must be within the valid population size. Environmental selection ranks only through the cutoff front, but ranks returned for selected individuals are complete. Use non_dominate_rank when you need ranks for every input point.

Backend modules also expose dominate_relation(x, y, cv_x=None, cv_y=None), which returns a dominance matrix between two sets of objectives. This function is not a top-level public export of evomo.operators.selection; access it through the backend module.

Evaluation and state helpers#

parse_evaluate(eval_out) returns (fitness, cv); pure objective tensors give cv=None.
at_least_2d handles single and batched inputs; see its generated return contract.
get_pareto_front and unique_rows_sorted filter and deduplicate results; consider dynamic output sizes when compiling.
load_pareto_front_from_file loads packaged constrained-problem reference fronts.
register_lazy_buffer supports state whose shape is known only after the first evaluation.

See the workflow guide and constraint guide for their corresponding semantics.