Skip to content

Records

Correctness fixes in HallOfFame, Logbook, History, and MultiStatistics.

ParetoFront crashed on a missing fitness

An empty front skipped a member without fitness (insert no-ops). Once the front held anyone, update compared ind.fitness and raised AttributeError, so later valid individuals were never seen.

Fix. Skip an individual that has no fitness attribute, matching HallOfFame.update.

Validator. tests/test_records/test_pareto_front.py::test_pareto_front_skips_individual_without_fitness

HallOfFame empty-archive bootstrap

If population[0] had no fitness, insert no-op'd and every later individual was skipped. HallOfFame(0) fell through to self[-1] and raised IndexError.

Fix. Return immediately when maxsize == 0. Bootstrap inserts the first individual that has a fitness attribute.

Validators.

  • tests/test_records/test_hall_of_fame.py::test_update_skips_no_fitness_bootstrap_and_keeps_later_members
  • tests/test_records/test_hall_of_fame.py::test_update_with_zero_maxsize_is_a_noop

HallOfFame discarded a strictly better clone

_has_similar ran after the fitness test and compared genotype only, so a noisier or later-better copy of an archived genome was dropped.

Fix. Skip a similar member that is equal or worse; if the newcomer is strictly better, replace the archive member.

Validator. tests/test_records/test_hall_of_fame.py::test_update_replaces_similar_member_when_fitness_is_better

remove desynchronized keys and items

An out-of-range index wrap-arounded keys then failed on items, leaving the two lists different lengths.

Fix. Normalize a negative index, then raise IndexError if it is still out of range.

Validator. tests/test_records/test_hall_of_fame.py::test_remove_out_of_range_keeps_keys_aligned

Logbook __delitem__ assumed chapters were row-aligned

A generation recorded without a chapter left that chapter shorter. del lb[i] then popped the wrong chapter row, or raised IndexError after the parent row was already gone.

Fix. Delete the chapter row that belongs to that parent record (match on gen), not chapter.pop(i).

Validator. tests/test_records/test_logbook.py::test_delitem_does_not_drop_unrelated_chapter_row

Slice delete had the same chapter desync

_delete_slice reused the parent's numeric index for every chapter.

Fix. Walk indexes high-to-low and call _delete_index, which matches chapter rows by gen and skips a missing one.

Validator. tests/test_records/test_logbook.py::test_delitem_slice_does_not_drop_unrelated_chapter_row

Repeated gen paired the wrong chapter row

Matching on gen alone popped the first chapter row with that value, not the occurrence that belonged to the deleted parent index.

Fix. Count parent rows with that gen at or after the deleted index and pop the matching occurrence.

Validator. tests/test_records/test_logbook.py::test_delitem_pairs_matching_repeat_generation

pop left chapter rows behind

__delitem__ removed the matching chapter row. pop only dropped the parent entry, so logbook.pop(i) and del logbook[i] diverged: chapters still held the deleted generation.

Fix. pop removes the chapter row that shares gen (same occurrence rule as __delitem__). __delitem__ now calls pop.

Validator. tests/test_records/test_logbook.py::test_pop_removes_matching_chapter_row

pop(-1) rewound the stream cursor

A raw negative index was compared to buff_index, so pop(-1) looked like “before the last streamed row” and rewound the cursor.

Fix. Convert index + len before comparing to buff_index.

Validator. tests/test_records/test_logbook.py::test_pop_negative_index_does_not_rewind_unstreamed_cursor

Stream assumed chapters were row-aligned

A generation recorded without a chapter left that chapter shorter than the parent. offsets used len(parent) instead of the chapter's own row count, so the chapter header was consumed as a data cell and later values landed on the wrong generation. stream then indexed past the rendered chapter lines (IndexError).

Fix. Render each chapter through a parent-aligned view. Rows pair by gen (the same occurrence rule as __delitem__). When there is no gen and the chapter is already the same length as the parent, pairing is positional. Missing chapter cells are blank.

Validators.

  • tests/test_records/test_logbook.py::test_stream_omitted_chapter_does_not_raise
  • tests/test_records/test_logbook.py::test_str_keeps_chapter_values_on_their_generation

to_json / from_json dropped nested chapters

Only one chapter level was walked. Two-level stats (chapter of chapters) vanished on a round-trip.

Fix. Recurse chapter.chapters.

Validator. tests/test_records/test_logbook.py::test_json_round_trip_restores_nested_chapters

History.update orphaned a whole batch

One individual without history_index made the batch's parent edges disappear, including for members that did have an index.

Fix. Collect history_index only from individuals that have the attribute.

Validator. tests/test_records/test_history.py::test_update_keeps_parents_when_one_individual_lacks_history_index

GridArchive NaN fitness replaced a finite elite

add required a valid single-objective fitness but did not require finite values. nan <= incumbent is False, so a NaN candidate replaced a finite elite and stats.qd_score became NaN. inf did the same. An empty cell would also store a non-finite elite.

Fix. Reject a candidate whose first weighted objective is not finite, the same way non-finite descriptors are rejected.

Validator. tests/test_records/test_grid_archive.py::test_add_rejects_non_finite_fitness

HallOfFame archived invalid and non-finite fitness

update only skipped a missing fitness attribute. A typical creator individual has the attribute before evaluation, so unevaluated members occupied hall-of-fame slots. Assigned NaN compared as incomparable under bisect_right, so a NaN key sorted to the front as if it were the best member. ParetoFront kept the same invalid and all-NaN individuals on the front.

Fix. Skip an individual whose fitness is missing, invalid, or non-finite, matching MAP-Elites add.

Validators.

  • tests/test_records/test_hall_of_fame.py::test_update_skips_invalid_fitness_and_keeps_later_members
  • tests/test_records/test_hall_of_fame.py::test_update_rejects_non_finite_fitness
  • tests/test_records/test_pareto_front.py::test_pareto_front_skips_invalid_and_non_finite_fitness

pop without gen left chapter rows behind

pop and __delitem__ pair chapter rows by gen. A record written without that key returned no match, so the chapter kept every row. Stream pairing already falls back to position when lengths match, so str then blanked the remaining cells.

Fix. When gen is missing and the chapter is the same length as the parent, delete (and render) the chapter row at the same index. A shorter chapter — a generation recorded without that chapter — is still left alone.

Validators.

  • tests/test_records/test_logbook.py::test_pop_without_generation_removes_positional_chapter_row
  • tests/test_records/test_logbook.py::test_pop_without_generation_keeps_shorter_chapter

GridArchive accepted non-finite ranges

parse_grid_config required low < high but not finite ends. (0, inf) built a grid that mapped every finite descriptor to cell 0. (-inf, 0) and NaN bounds crashed later in descriptor_to_index with int(nan).

Fix. Reject a range whose ends are not finite.

Validator. tests/test_records/test_grid_archive_errors.py::test_non_finite_ranges_raise

MultiStatistics.compile exhausted a generator

Each chapter iterated data. A one-shot iterable (generator, map, zip) was empty for every chapter after the first.

Fix. data = list(data) once.

Validator. tests/test_records/test_statistics.py::TestStatistics::test_multi_statistics_compile_consumes_generator_once

HallOfFame.insert archived invalid and non-finite fitness

update already skips missing, invalid, and non-finite fitness. insert only required a fitness attribute, so a direct insert or from_json restore could archive wvalues=() or NaN. Those keys then sorted as if they were the best member.

Fix. insert uses the same comparable-fitness guard as update. ParetoFront.insert and JSON restore go through that path.

Validators.

  • tests/test_records/test_hall_of_fame_insert.py::test_insert_skips_invalid_and_non_finite_fitness
  • tests/test_records/test_hall_of_fame_insert.py::test_pareto_insert_skips_invalid_and_non_finite_fitness
  • tests/test_records/test_hall_of_fame_insert.py::test_from_json_does_not_restore_non_finite_fitness

clear left chapter rows and the stream cursor

pop and __delitem__ keep chapters and buff_index aligned. list.clear does not go through those methods. After clear(), chapters still held every generation. After a stream, clear(), and a new record, stream started at the stale cursor and returned no rows.

Fix. clear deletes every parent row through __delitem__ (the same chapter and cursor rules as del logbook[:]).

Validators.

  • tests/test_records/test_logbook_edges.py::test_clear_removes_chapter_rows
  • tests/test_records/test_logbook_edges.py::test_clear_after_stream_does_not_drop_new_rows