94g: threshold objective for collapse_global (fail-count, not continuous)

Add objective="quality"|"threshold" to collapse_global and make threshold
the finish-time default. The continuous-quality objective maximises
sum(usage_quality*area), which can trade one leaf just over the 0.1 fail
threshold for another just under (a fail SHUFFLE). The threshold objective
maximises the COUNT of passing size/width/proportion factors directly, with
continuous fit only as a tiebreak. A satisfied adjacency and a passing factor
share one weight (_COLLAPSE_FAIL_W) so both fail classes are minimised jointly.

Sweep over 6 harbor-house evolved layouts (total fails, base 195):
  adj_off/quality 192  adj_on/quality 185  adj_off/thresh 181  adj_on/thresh 172
adj_on/threshold is monotone across all 6 (never worse than baseline), so it
is the new default. Residual on the best layout (15→13) is the building-level
"no outside public access" constraint, outside the per-leaf model.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01M8566xAxTnwtJTkpXjYNZm
This commit is contained in:
Bruno Postle 2026-07-18 08:37:02 +01:00
parent d52cce6863
commit da18ef744e
2 changed files with 63 additions and 30 deletions

File diff suppressed because one or more lines are too long

View file

@ -327,9 +327,18 @@ class Fitness:
# finite (Hungarian cannot take -inf) yet far below any real value, so the # finite (Hungarian cannot take -inf) yet far below any real value, so the
# optimal matching never uses a level-mismatched pair unless it is forced. # optimal matching never uses a level-mismatched pair unless it is forced.
_COLLAPSE_FORBID = -1e12 _COLLAPSE_FORBID = -1e12
# Weight of one avoided fail (a satisfied adjacency or a passing
# size/width/proportion factor) in the collapse objective — far above the
# continuous quality span (~max area) so fail count dominates and raw
# quality only breaks ties; far below the forbid penalty so level holds.
_COLLAPSE_FAIL_W = 1e6
def collapse_global( def collapse_global(
self, root: Node, adjacency: bool = True, iters: int = 6 self,
root: Node,
adjacency: bool = True,
objective: str = "threshold",
iters: int = 6,
) -> None: ) -> None:
"""Finish-time GLOBAL cell->room collapse (homemaker-py-94g): relabel """Finish-time GLOBAL cell->room collapse (homemaker-py-94g): relabel
every inside-room leaf across the whole building to the required room it every inside-room leaf across the whole building to the required room it
@ -354,8 +363,17 @@ class Fitness:
a labelling relaxation: warm-started from the evolved labels, each pass is a labelling relaxation: warm-started from the evolved labels, each pass is
a linear assignment over quality + adjacency-bonus computed from the a linear assignment over quality + adjacency-bonus computed from the
previous pass, iterated to a fixpoint (Jacobi/WFC-style). Maximising previous pass, iterated to a fixpoint (Jacobi/WFC-style). Maximising
satisfied adjacencies minimises adjacency fails. The base per-leaf value satisfied adjacencies minimises adjacency fails.
is the separable sum(usage_quality * area) collapse_superposition uses.
OBJECTIVE selects the per-leaf base value: ``"quality"`` maximises the
separable continuous fit sum(usage_quality * area) collapse_superposition
uses; ``"threshold"`` maximises the COUNT of size/width/proportion factors
that PASS (>= FAIL_THRESHOLD), with continuous fit only as a tiebreak.
Continuous quality can trade one leaf just over threshold for another just
under (a fail SHUFFLE); the threshold objective optimises the fail count
directly. Under both, a satisfied adjacency and a passing factor carry the
same weight (_COLLAPSE_FAIL_W = one avoided fail), so the collapse
minimises (adjacency + size/width/proportion) fails jointly.
One-shot finish-time pass on a committed layout, not a per-eval re-type.""" One-shot finish-time pass on a committed layout, not a per-eval re-type."""
prog = self._programme or {} prog = self._programme or {}
@ -375,18 +393,35 @@ class Fitness:
return return
forbid = self._COLLAPSE_FORBID forbid = self._COLLAPSE_FORBID
fail_w = self._COLLAPSE_FAIL_W
levels_of = [dom_mod.level_of(lf) for lf in supply] levels_of = [dom_mod.level_of(lf) for lf in supply]
areas = [geometry.area(lf) for lf in supply] areas = [geometry.area(lf) for lf in supply]
# Base (separable) quality: usage fit x area, or forbid on level mismatch. # Base per-cell value: forbid on level mismatch, else the separable fit.
# In "threshold" mode add fail_w per passing size/width/proportion factor
# so the matching maximises passes first, continuous fit only as tiebreak.
base: list[list[float]] = [] base: list[list[float]] = []
for i, lf in enumerate(supply): for i, lf in enumerate(supply):
orig = lf.type
row = [] row = []
for code in slots: for code in slots:
req = prog[code] req = prog[code]
if req.level is not None and req.level != levels_of[i]: if req.level is not None and req.level != levels_of[i]:
row.append(forbid) row.append(forbid)
else: continue
row.append(self._usage_quality(lf, code) * areas[i]) lf.type = code
qs = self.quality_size(lf)
qw = self.quality_width(lf)
qp = self.quality_proportion(lf)
val = qs * qw * qp * areas[i]
if objective == "threshold":
passes = (
(qs >= FAIL_THRESHOLD)
+ (qw >= FAIL_THRESHOLD)
+ (qp >= FAIL_THRESHOLD)
)
val += fail_w * passes
row.append(val)
lf.type = orig
base.append(row) base.append(row)
if not adjacency: if not adjacency:
@ -396,13 +431,11 @@ class Fitness:
return return
# Adjacency relaxation. Build the pre-merge base graph once (fixed # Adjacency relaxation. Build the pre-merge base graph once (fixed
# geometry) and weight a satisfied adjacency above any quality span so # geometry). A satisfied adjacency is worth fail_w — one avoided fail,
# avoiding an adjacency fail always outranks a size/width/proportion gain. # the same unit as a passing factor — so both are minimised jointly.
from . import graph as graph_mod from . import graph as graph_mod
graphs = graph_mod.build_graphs(root, self.conf("door_width") or 1.2) graphs = graph_mod.build_graphs(root, self.conf("door_width") or 1.2)
adj_w = (max(areas) if areas else 1.0) + 1.0
# Distinct codes among the slots, with their required adjacency lists.
code_adj = {code: prog[code].adjacency for code in set(slots)} code_adj = {code: prog[code].adjacency for code in set(slots)}
prev_labels: list[str | None] = None # type: ignore[assignment] prev_labels: list[str | None] = None # type: ignore[assignment]
@ -418,7 +451,7 @@ class Fitness:
for ac in code_adj[code] for ac in code_adj[code]
if graph_mod.has_adjacency(lf, ac, G) if graph_mod.has_adjacency(lf, ac, G)
) )
quality[i][j] += adj_w * sat quality[i][j] += fail_w * sat
assign = self._best_assignment(quality) assign = self._best_assignment(quality)
new_labels: list[str | None] = [lf.type for lf in supply] new_labels: list[str | None] = [lf.type for lf in supply]
for r, c in assign: for r, c in assign: