DEAP/deap issues ported into deap-er¶
Local notes for the DEAP/deap issues implemented in this repo on 2026-09-06. Source tracker: https://github.com/DEAP/deap/issues.
These are ports of the agreed Should implement and Could implement items. They are not upstream DEAP patches. Default Fitness comparison, sel_nsga_2 crowding, and docs/README were left alone.
Operator reproductions for #655, #740, #527, #472, and #755 live in 1db151f with the public exports.
#655 — mutPolynomialBounded returns NaN¶
Issue. DEAP#655. Polynomial mutation computed delta from a gene already outside [low, up]. The base of ** mut_pow went negative, Python produced a complex or nan, and that value was written back onto the individual.
Why it was an issue. NSGA-II and other bounded real-coded GAs start or drift genes outside the documented box (tight per-gene bounds, earlier unbounded variation). A single out-of-box gene poisoned fitness and the rest of the run. eta <= 0 also hits 1/(eta+1).
What we did. mut_polynomial_bounded now requires eta > 0, skips a gene when xu <= xl, and clamps the gene into [xl, xu] before any power. The Deb formula and the final clamp are unchanged, so in-box golden cases stay the same.
Commit. b43d440eb5d77665f88bb448c8eeceb95846024a
#740 — cxSimulatedBinaryBounded raises on complex math¶
Issue. DEAP#740. Bounded SBX built beta from parent-to-bound gaps. Out-of-box parents (or a bad eta) made beta negative, alpha complex, and <= between float and complex raised TypeError.
Why it was an issue. Same class of bug as #655: the operator assumed every parent gene was already inside the box. Real populations are not. Empty intervals (xu == xl) also divide by zero.
What we did. Same require_positive_eta helper. Skip xu <= xl. Clamp both parents into [xl, xu] before the Deb powers; if they coincide after the clamp, leave the locus unchanged. In-box numeric results are unchanged. A skipped 50% draw can still leave an out-of-box gene (clamp is only on the power path).
Commit. b43d440eb5d77665f88bb448c8eeceb95846024a
#527 — Bounded blend crossover¶
Issue. DEAP#527. cxBlend has no box. Design problems that already used cxSimulatedBinaryBounded wanted the same low/up on blend.
Why it was an issue. Blend draws past the parents by alpha. Without a clamp, children leave a constrained search space that the rest of the toolbox treats as a box.
What we did. Added cx_blend_bounded(ind1, ind2, alpha, low, up): same gamma as cx_blend, then clamp each child gene. Empty intervals are skipped. Exported from operators.py / tools.
Commits. Operator: b43d440eb5d77665f88bb448c8eeceb95846024a. Public export and tests: 1db151fcdc408fae6f9bb1d86c17b1fa87d6dc25.
#472 — PMX only works for {0..n-1}¶
Issue. DEAP#472. cxPartialyMatched (and uniform PMX) built position maps as p1[ind1[i]] = i. That only works when alleles are the indices.
Why it was an issue. TSP-style and named permutations (['a','b','c'], city IDs) IndexError or silently write the wrong genes. cx_ordered had the same allele-as-index hole array (holes1[ind2[i]]). One root cause, three operators.
What we did. _allele_maps builds dict allele→index maps and raises ValueError on duplicates or mismatched gene sets. match uses those dicts. cx_ordered uses membership sets for the kept slice. {0..n-1} still works.
Commit. c25412af210f9d1074373a8b35a4b845c788fd9b (tests in 1db151fcdc408fae6f9bb1d86c17b1fa87d6dc25)
#755 — Heterogeneous / per-gene mutation¶
Issue. DEAP#755. Every stock mutator treats every gene the same (flip all bits, Gaussian all floats). Mixed individuals (bit + int range + choice) had no operator.
Why it was an issue. Mixed-representation GAs are common in design/config search. Users had to write a one-off mutator for every encoding.
What we did. mut_heterogeneous(individual, mutators, mut_prob): mutators[i](value) -> value, one callable per gene. Raises if the lengths differ. Lives in its own module so mut_various.py stays under the file-length cap. Exported from tools.
Commit. 1db151fcdc408fae6f9bb1d86c17b1fa87d6dc25
#500 — CMA lower and upper bounds¶
Issue. DEAP#500. Strategy.generate sampled an unbounded Gaussian. The only documented workaround was a penalty on evaluate, or switching to the standalone cma package.
Why it was an issue. Most engineering CMA runs are box-constrained. Unbounded samples waste evaluations and confuse anyone coming from bounded SBX/poly.
What we did. low / up (scalar or length-dim) and bound_mode of "clip" (default) or "resample" on every CMA strategy. Applied to the sampled vector before ind_init. Resample rejects the whole vector (truncated MVN, not per-gene redraw). After resample_limit (default 100) failed draws, that offspring is clipped so generate always returns lamb individuals. StrategyMultiObjective.generate is wired separately because it does not call sample_offspring. A later compute_params() without bound kwargs keeps existing bounds. Docstrings state both modes are constraint-handling approximations: the CMA update then treats the repaired point as the sample.
Commit. 4465e9e89ebc0dc981a165eec26cd7dd406f3bc6
#383 — Weighted GP primitive sampling¶
Issue. DEAP#383. generate picked primitives uniformly. Users wanted add_primitive(..., weight=4) so some functions appear more often.
Why it was an issue. Uniform sampling over {add, sin, tan, ...} over-represents rare operators and under-represents the arithmetic core of symbolic regression.
What we did. Primitive.weight defaults to 1.0 so existing trees stay equal. add_primitive(..., weight=1.0) on typed and untyped sets; weight <= 0 is rejected. choose_weighted uses rng.choice when all candidate weights are equal (same RNG stream as before). Only unequal weights walk a cumulative sum with rng.random(). Same helper in mut_node_replacement and mut_insert. Terminals stay uniform. No rng.choices (rng.py is at the 270-line hard cap).
Commit. e4b9d05984595d4cf990233d866c92bd1d4619a3
#644 — Zero-arity callable terminals format as name not name()¶
Issue. DEAP#644. A 0-arity function registered as a terminal stringified as no_input_func_1. compile / eval then looked up the function object instead of calling it.
Why it was an issue. The Python compile path is eval of str(tree). Without (), a 0-arity terminal is a name, not a call, so the compiled program is wrong or raises.
What we did. Terminal.call_zero. add_terminal sets it when the registered context value is a named callable. format() emits name(). ARG terminals, numbers, and True/False stay unwrapped. from_string still tokenizes on (); tokens remain no_input_func_1.
Commit. e4b9d05984595d4cf990233d866c92bd1d4619a3
#24 — Infix pretty-printer for GP trees¶
Issue. DEAP#24. Trees only printed as prefix add(x, y). Symbolic-regression users wanted x + y.
Why it was an issue. Prefix dumps are hard to read in logs and papers. This is display-only; it is not a second compiler.
What we did. New module deap_er/private/programming/infix.py with tree_to_infix. Maps add/sub/mul/div/neg and the v* numpy aliases to infix; unknown primitives stay name(args). Exported from gp.py. Not added to compilers.py (already near the line cap).
Commit. e4b9d05984595d4cf990233d866c92bd1d4619a3
#321 — Crowding distance on wvalues¶
Issue. DEAP#321. assignCrowdingDistance crowds on fitness.values. The request was to crowd on wvalues so scale/sign of weights affect spacing.
Why it was an issue. Deb’s NSGA-II uses raw objectives. Some users want weighted space when objective magnitudes differ. Changing the default would change every NSGA-II run.
What we did. assign_crowding_dist(..., *, use_weights=False). Default stays values. sel_nsga_2 is unchanged. Note: for nonzero per-objective scales the Deb formula is affine-invariant, so use_weights=True only changes distances when a weight is 0 (that objective is skipped) or the caller has customized wvalues.
Commit. b4abadb2b22a60e6bb93b0b9d079bdb510bd7365
#641 / #247 — DCD tournament requires k divisible by 4¶
Issue. DEAP#641, DEAP#247. selTournamentDCD walks two shuffled copies in steps of 4. k=1 or k=9 raised IndexError in DEAP; deap-er already raised ValueError for k % 4 != 0.
Why it was an issue. NSGA-II examples use k == len(pop) and even sizes. Anyone who wanted len(pop)-1 or an odd k could not use DCD selection.
What we did. Keep the original paired-shuffle path when sel_count % 4 == 0 (same results as before). For other k in 1..len(individuals), run pairwise DCD tournaments on shuffled copies until k winners. sel_count > len is still ValueError. sel_count <= 0 returns [].
Commit. b4abadb2b22a60e6bb93b0b9d079bdb510bd7365
#720 — MultiStatistics chapters need different functions¶
Issue. DEAP#720. MultiStatistics.register forwarded the same function to every chapter. Fitness min/max and a GP terminal-count dict cannot share numpy.mean.
Why it was an issue. The workaround was to register on each inner Statistics before wrapping, or to change every ea_* signature to take several stats objects.
What we did. MultiStatistics.register(..., chapters=None). None still registers on every chapter. A str or iterable names a subset. chapters is keyword-only so it is not bound into the statistic function.
Commit. 3cb9555b62a78c18f69ad9eed290bbc90237bea5
#694 — Logbook.stream should print a header when empty¶
Issue. DEAP#694. An empty logbook’s stream() printed “empty” instead of the column banner. Generation 0 can take a long time; users wanted the table header immediately.
Why it was an issue. build_header also assumed there were data rows (max over an empty str_matrix crashed). After we printed a header on empty, a naive start_index == 0 reprint doubled the banner. Setting header_streamed on any empty stream() then dropped the banner for a book that had no header set.
What we did. Empty logbook with a header prints that header. header_streamed is set only when a header was actually emitted (self.header or the book already has rows), so a bare Logbook().stream() still says empty and the first real stream() still prints columns. Header-only sizing uses column-name widths.
Commit. 3cb9555b62a78c18f69ad9eed290bbc90237bea5
#121 — Serialize logbook to JSON¶
Issue. DEAP#121. json.dumps(logbook) lost chapters and numpy scalars.
Why it was an issue. Downstream tools and notebooks want a portable table, not a dill checkpoint.
What we did. Logbook.to_json / from_json serialize entries, chapters, and header. Numpy scalars become float/int via .item(); leftover non-JSON values are stringified. Round-trip restores chapters.
Commit. 3cb9555b62a78c18f69ad9eed290bbc90237bea5
#350 — Duplicate-count statistic¶
Issue. DEAP#350. Users wanted the number of twin individuals per generation as a Statistics field.
Why it was an issue. Variety collapse is invisible if you only log mean fitness.
What we did. duplicate_count(population, key=None) in metrics.py (identity key by default): len(population) minus distinct keys. Users register it on Statistics. Already star-exported via tools.
Commit. 3cb9555b62a78c18f69ad9eed290bbc90237bea5
#426 — Per-generation wall time¶
Issue. DEAP#426. Papers want a time-per-generation curve. The algorithms did not record duration.
Why it was an issue. Users wrapped ea_* or timed the whole run and lost per-gen resolution.
What we did. Optional log_time=False on all four ea_* functions. time.perf_counter() around each generation, including gen 0 evaluation. Writes duration into the record. "duration" is added to the printed header only when log_time is True, so the default table stays gen, nevals.
Commit. 3cb9555b62a78c18f69ad9eed290bbc90237bea5
#750 — Logger instead of print¶
Issue. DEAP#750. verbose used print. Inside Docker or a service, stdout is often not the log sink.
Why it was an issue. print also interacts badly with a destructive logbook.stream: a second read is empty, so print plus logger.info(logbook.stream) would log nothing.
What we did. Optional logger: logging.Logger | None = None on the four ea_* functions. When verbose, call stream() once; print that string if logger is None, else logger.info(...).
Commit. 3cb9555b62a78c18f69ad9eed290bbc90237bea5
#735 — Pareto front every generation¶
Issue. DEAP#735. eaMuPlusLambda plus a ParetoFront HoF only yields the cumulative front at the end. Users wanted the front of each generation.
Why it was an issue. A HoF/ParetoFront is cumulative. Deepcopying it does not give “the front of generation t”.
What we did. Optional fronts: list | None = None on all four ea_* functions. After the HoF update, if fronts is a list, append a new ParetoFront built from the current population (that generation’s survivors, not the HoF). Length is gens+1 for algorithms with a gen-0 record, gens for ea_generate_update.
Commit. 3cb9555b62a78c18f69ad9eed290bbc90237bea5