Hacktoberfest 2026: le issue che i maintainer hanno segnato per ottobre, aperte e adatte ai principianti. Sfoglia le issue Hacktoberfest

An altpenalty compared against the constant 0 (altpenalty X > 0, X >= 0, 0 < X) is silently ignored, so a failed constraint is charged weight times its primary violation instead

Aperta Adatta ai principianti
  • Qualcosa non va?
#889 0 commenti 0 reazioni 0 assegnatari Vedi su GitHub

I maintainer di solito rispondono entro 1 giorno

Nessuno ha ancora preso questa issue.

Valutazione

Difficoltà
2/5
Tempo stimato
1-3 ore
Idoneità per principianti
80/100
Tipo di issue
Bug
Chiarezza
Specificata chiaramente
Stato di attività
Attiva
Stack tecnologico
python
Ambito
data

Direzione di ricerca

Leggere la normalizzazione di altpenalty in pybnf/constraint.py:338-347, quindi esaminare i controlli in get_static_penalty alla riga 689 e in _static_penalty_gradient alla riga 889. Riprodurre il problema con l’esempio fornito obs.prop e sim.gdat e verificare che sia il percorso del penalty sia quello del gradient gestiscano una costante zero. Il lavoro è completato quando i casi interessati producono i penalty corretti elencati e i controlli esistenti rimangono invariati.

Scritto dal modello di indicizzazione a partire dal testo della issue.

Descrizione

bug silent-incorrectness

load_constraint_file turns a numeric side of the altpenalty inequality into a float (pybnf/constraint.py:113-117). Constraint.init then rewrites a >/>= altpenalty as < by swapping its sides (pybnf/constraint.py:338-347). After both steps, altpenalty X > 0, X >= 0, 0 < X and 0 <= X all store alt1 = 0.0, alt2 = 'X'. get_static_penalty uses if self.alt1: (pybnf/constraint.py:689) to decide whether an altpenalty is present. Because 0.0 is falsy, the altpenalty branch is skipped and the penalty becomes weight times the primary inequality's violation. _static_penalty_gradient runs the same test (pybnf/constraint.py:889), so the gradient matches the wrong objective and does not expose the problem. find_keys accepts a float alt1 without complaint (pybnf/constraint.py:421-422, where get_key returns None). No error, warning or log line appears. altpenalty X < 0 is not affected, because there the 0 ends up in alt2.

Failure scenario

docs/config.rst:251 defines the altpenalty rule. When the primary inequality fails, the penalty is weight * max(0, alt1 - alt2). If the altpenalty inequality holds, the penalty is weight * min, or 0 when no min is set. Take A < 1 at 1 weight 10 altpenalty X > 0 with A = 5:

  • X = -3: the penalty should be 10 * max(0, 0 - (-3)) = 30. PyBNF returns 40, which is 10 times the primary violation 5 - 1.
  • X = +3: the altpenalty inequality holds, so the penalty should be 0. PyBNF returns 40.
  • X = +3 with min 1: the penalty should be 10. PyBNF returns 40.

The error runs in both directions. A parameter set that satisfies the continuous proxy is still charged, and the penalty follows A instead of X, which defeats the purpose of the substitution.

Reproduction

Run from any empty directory with PyBNF importable:

from pybnf import data, constraint

lines = ['A < 1 at 1 weight 10 altpenalty X > 0',
         'A < 1 at 1 weight 10 altpenalty 0 < X',
         'A < 1 at 1 weight 10 altpenalty X >= 0',
         'A < 1 always weight 10 altpenalty X > 0',
         'A < 1 at 1 weight 10 altpenalty X > 0 min 1',
         'A < 1 at 1 weight 10 altpenalty X > 1e-300 min 1',  # control
         'A < 1 at 1 weight 10 altpenalty X > 0.5']           # control
with open('obs.prop', 'w') as f:
    f.write('\n'.join(lines) + '\n')
for x in (-3, 3):
    with open('sim.gdat', 'w') as f:
        f.write(f'# time A X\n0 5 {x}\n1 5 {x}\n')
    d = data.Data()
    d.load_data('sim.gdat')
    cs = constraint.ConstraintSet('model', 'obs')
    cs.load_constraint_file('obs.prop')
    for line, c in zip(lines, cs.constraints):
        print(f'X={x:+d}  {line:50s} alt1={c.alt1!r:7} penalty={c.penalty({"model": {"obs": d}})}')
constraint X PyBNF correct
altpenalty X > 0 -3 40.0 30
altpenalty 0 < X -3 40.0 30
altpenalty X >= 0 -3 40.0 30
always ... altpenalty X > 0 -3 40.0 30
altpenalty X > 0 min 1 -3 40.0 30
altpenalty X > 1e-300 min 1 (control) -3 30.0 30
altpenalty X > 0.5 (control) -3 35.0 35
altpenalty X > 0 +3 40.0 0
altpenalty 0 < X +3 40.0 0
altpenalty X >= 0 +3 40.0 0
always ... altpenalty X > 0 +3 40.0 0
altpenalty X > 0 min 1 +3 40.0 10
altpenalty X > 1e-300 min 1 (control) +3 10.0 10
altpenalty X > 0.5 (control) +3 0.0 0

Every affected line prints alt1=0.0. Moving the constant only to 1e-300 gives the correct value.

Reachability

This affects any .prop constraint file (qualitative data read through exp_file / data:) with a weighted constraint whose altpenalty has 0 as its constant side after normalization: X > 0, X >= 0, 0 < X or 0 <= X. It applies to at, between, always and once constraints. The altpenalty grammar accepts a number on either side (pybnf/constraint.py:249-251, :264), and X > 0 is the natural way to say that the continuous proxy should be positive. Nothing upstream rejects the input and nothing downstream corrects it. Gradient-based fits get a gradient consistent with the wrong penalty.

Where

pybnf/constraint.py:689 pybnf/constraint.py:889 pybnf/constraint.py:338-347

The obvious fix is if self.alt1 is not None: at both test sites.

Related: #890, #887.

Found in a whole-codebase audit for silently wrong results (2026-09-23); the reproduction above was re-run independently of the original finding.

Lingua principale
Python
Stelle
25
Fork
25
Merge medio
3h 32m
PR unite (30g)
96

Preparare l'ambiente

Come iniziare

  1. Leggi tutta la issue e poi la guida ai contributi del progetto.
  2. Commenta sulla issue per dire che te ne occupi tu — evita che due persone facciano lo stesso lavoro.
  3. Fai un fork del repository e lavora su un branch.
  4. Apri una pull request che faccia riferimento al numero della issue.

Altre issue di lanl/PyBNF

Tutte le issue di lanl/PyBNF

Issue simili

Altre issue su Python

Ricevi le nuove issue nella tua casella

Un breve riepilogo di issue GitHub adatte ai principianti.