Bruno Postle
3b3eef68a7
Ramp the leaf-share grain down within one continuous run (e.g. 4->3->2->off), carrying the whole population across each step — graduated non-convexity over the single hard sharing->off transition of the §15 finish. - operators.unfold_shared_leaves(above=cap): unfold only leaves whose share exceeds the new grain cap, leaving smaller-share leaves collapsed for the next step. above=1 (default) keeps the full-unfold §15 behaviour. - driver: max_share override threaded through _overrides_for/_fitness_for/ _evaluate so a phase can rebuild the evaluator at a lower leaf_share_max cap; search(seed_pop=) evaluates an explicit initial population so a phase hands its whole population to the next instead of restarting from a single best. - driver.search_annealed: one phase per descending grain then a de-share polish; unfold-above-cap between steps; cumulative accounting + grain-tagged history; honest canonical best (byte-for-byte verified vs homemaker-fitness). - evolve: --anneal-grain LADDER CLI (self-finishing; §15 finish not applied). 8iv settled the primitive (grid unfold beat the circulation-aware slice), so the ramp reuses the plain balanced-grid unfold at every step. Tests: unfold above-cap selectivity, seed_pop seeding, search_annealed phase stitching / honest finish / degenerate-ladder fallback. 258 pass. DESIGN §16. Head-to-head A/B on harbor-house still to run; verdict pending (issue open). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01M8566xAxTnwtJTkpXjYNZm
362 lines
17 KiB
Python
362 lines
17 KiB
Python
"""Driver tests with a faked inner loop (no oracle, no perl)."""
|
|
|
|
from pathlib import Path
|
|
|
|
import numpy as np
|
|
import pytest
|
|
|
|
from homemaker_layout import dom, driver, innerloop, solver
|
|
|
|
CORPUS = Path(__file__).parent.parent / "examples" / "programme-house"
|
|
SEED_FILE = CORPUS / "c964435454c459f86c3ed9a5a7621132.dom"
|
|
INIT_FILE = CORPUS / "init.dom"
|
|
|
|
pytestmark = pytest.mark.skipif(not CORPUS.is_dir(), reason="Corpus not available")
|
|
|
|
|
|
def test_free_with_keys_aligns_with_free_branches():
|
|
for f in sorted(CORPUS.glob("*.dom")):
|
|
root = dom.load(str(f))
|
|
assert [b for _, b in innerloop.free_with_keys(root)] == solver.free_branches(root), f.name
|
|
|
|
|
|
@pytest.fixture
|
|
def fake_inner(monkeypatch):
|
|
"""Deterministic fake fitness: rewards leaf count up to 12; consumes the
|
|
full budget; applies a recognisable ratio so Lamarckian write-back is
|
|
observable."""
|
|
calls = []
|
|
|
|
def fake_optimise(root, programme_dir, x0=None, budget=200, urb_root=None, **kw):
|
|
n_leaves = sum(len(lvl.leaves()) for lvl in dom.levels(root))
|
|
fitness = 1.0 / (1.0 + abs(12 - n_leaves)) + 1e-6 * len(calls)
|
|
calls.append({"budget": budget, "x0": x0, "kw": kw})
|
|
for _, b in innerloop.free_with_keys(root):
|
|
b.division = [0.25, 0.25]
|
|
return innerloop.Result(
|
|
x=np.array([0.25]), fitness=fitness, n_fails=0, fail_lines=(),
|
|
x0_fitness=fitness / 2, x0_n_fails=1, n_evals=budget, n_oracle_calls=1,
|
|
)
|
|
|
|
monkeypatch.setattr(innerloop, "optimise", fake_optimise)
|
|
return calls
|
|
|
|
|
|
def test_search_respects_budget_and_logs(fake_inner):
|
|
seed_root = dom.load(str(SEED_FILE))
|
|
lines = []
|
|
r = driver.search(seed_root, CORPUS, budget=1000, pop_size=4,
|
|
child_budget=80, seed_budget=120, seed=1, log=lines.append)
|
|
# budget accounting: seed (120) + children (80 each), stop at >= 1000
|
|
assert r.n_evals >= 1000
|
|
assert r.n_evals == 120 + 80 * ((r.n_evals - 120) // 80)
|
|
assert r.n_evals - 1000 < 80
|
|
assert r.n_topologies == 1 + (r.n_evals - 120) // 80
|
|
assert lines, "improvements must be logged"
|
|
# history monotone in evals and fitness
|
|
evs = [h[0] for h in r.history]
|
|
fits = [h[1] for h in r.history]
|
|
assert evs == sorted(evs)
|
|
assert fits == sorted(fits)
|
|
assert r.best.fitness == max(fits)
|
|
assert len(r.population) <= 4
|
|
# Lamarckian write-back observable in the best individual
|
|
assert all(b.division == [0.25, 0.25] for _, b in innerloop.free_with_keys(r.best.root))
|
|
|
|
|
|
def test_search_children_warm_start_and_local_sigma(fake_inner):
|
|
seed_root = dom.load(str(SEED_FILE))
|
|
driver.search(seed_root, CORPUS, budget=500, pop_size=4,
|
|
child_budget=60, seed_budget=100, seed=0)
|
|
seed_call, child_calls = fake_inner[0], fake_inner[1:]
|
|
assert seed_call["x0"] is None and seed_call["budget"] == 100
|
|
assert child_calls
|
|
for c in child_calls:
|
|
assert c["budget"] == 60
|
|
assert c["x0"] is not None # warm-started
|
|
# inherited cuts carry the parent's written-back ratios
|
|
assert np.isin(c["x0"], [0.25, 0.5]).all()
|
|
assert "sigmas" not in c["kw"] # NM inner loop takes no sigmas
|
|
|
|
|
|
def test_best_root_dumps_valid_dom(fake_inner, tmp_path):
|
|
seed_root = dom.load(str(SEED_FILE))
|
|
r = driver.search(seed_root, CORPUS, budget=400, pop_size=3,
|
|
child_budget=60, seed_budget=100, seed=2)
|
|
out = tmp_path / "best.dom"
|
|
dom.dump(r.best.root, str(out))
|
|
reloaded = dom.load(str(out))
|
|
assert sum(len(lvl.leaves()) for lvl in dom.levels(reloaded)) == \
|
|
sum(len(lvl.leaves()) for lvl in dom.levels(r.best.root))
|
|
|
|
|
|
def test_bootstrap_cold_start(fake_inner):
|
|
"""Bootstrap auto-triggers from a bare undivided plot and fills the
|
|
population with pop_size diverse random topologies before the main loop."""
|
|
init_root = dom.load(str(INIT_FILE))
|
|
assert not init_root.divided, "init.dom should be an undivided bare plot"
|
|
|
|
pop_size = 4
|
|
child_budget = 60
|
|
budget = 500
|
|
r = driver.search(init_root, CORPUS, budget=budget, pop_size=pop_size,
|
|
child_budget=child_budget, seed_budget=100, seed=7)
|
|
|
|
# All evaluations use child_budget (no seed_budget call)
|
|
assert r.n_evals % child_budget == 0
|
|
assert r.n_evals >= budget
|
|
assert r.n_evals - budget < child_budget
|
|
# Every topology (bootstrap + main loop) is counted
|
|
assert r.n_topologies == r.n_evals // child_budget
|
|
# Population is full
|
|
assert len(r.population) == pop_size
|
|
# Bootstrap individuals all had x0=None (cold starts)
|
|
assert all(c["x0"] is None for c in fake_inner[:pop_size])
|
|
# Bootstrap uses exploratory sigma schedule (inner_kw={}, no sigmas override)
|
|
assert all("sigmas" not in c["kw"] for c in fake_inner[:pop_size])
|
|
# Main loop children are warm-started
|
|
main_calls = fake_inner[pop_size:]
|
|
assert main_calls # at least one main-loop child
|
|
assert all(c["x0"] is not None for c in main_calls)
|
|
|
|
|
|
def test_bootstrap_disabled_for_divided_seed(fake_inner):
|
|
"""A divided seed (warm start) auto-selects the legacy single-seed path."""
|
|
seed_root = dom.load(str(SEED_FILE))
|
|
assert seed_root.divided
|
|
|
|
r = driver.search(seed_root, CORPUS, budget=500, pop_size=4,
|
|
child_budget=60, seed_budget=100, seed=0)
|
|
|
|
# First call is the seed evaluated at seed_budget
|
|
assert fake_inner[0]["budget"] == 100
|
|
assert fake_inner[0]["x0"] is None
|
|
# Remaining are warm-started children at child_budget
|
|
assert all(c["budget"] == 60 for c in fake_inner[1:])
|
|
|
|
|
|
def test_random_topology_leaf_count():
|
|
"""random_topology produces a topology with at least n_leaves leaves."""
|
|
import numpy as np
|
|
init_root = dom.load(str(INIT_FILE))
|
|
rng = np.random.default_rng(0)
|
|
types = ["b1", "b2", "l1", "t1", "t2", "t3", "C", "O"]
|
|
for n in (3, 5, 7, 10):
|
|
topo = driver.random_topology(init_root, n, rng, types)
|
|
n_leaves = sum(len(lvl.leaves()) for lvl in dom.levels(topo))
|
|
assert n_leaves >= n
|
|
assert n_leaves <= n + 1 # mutate_divide adds exactly one leaf per call
|
|
|
|
|
|
def test_niche_by_signature_keeps_distinct_topologies(fake_inner):
|
|
"""§11.5: niching admits at most one individual per topology signature, so
|
|
the population is structurally distinct and diversity is reported."""
|
|
from homemaker_layout import genome
|
|
init_root = dom.load(str(INIT_FILE))
|
|
r = driver.search(init_root, CORPUS, budget=2000, pop_size=6,
|
|
child_budget=60, seed=3, niche_by_signature=True)
|
|
sigs = [genome.signature(p.root) for p in r.population]
|
|
assert len(sigs) == len(set(sigs)), "population must be one-per-topology"
|
|
assert r.n_distinct_signatures >= len(r.population)
|
|
assert r.diversity_history # recorded on each improvement
|
|
|
|
|
|
def test_restart_keeps_elite_and_counts(monkeypatch):
|
|
"""§11.5: a stagnation restart fires, is counted, and preserves the best."""
|
|
# Saturating fake (no monotone tiebreaker, unlike `fake_inner`): fitness
|
|
# peaks at 12 leaves and plateaus, so the best stalls and restarts trigger.
|
|
def fake_optimise(root, programme_dir, x0=None, budget=200, urb_root=None, **kw):
|
|
n_leaves = sum(len(lvl.leaves()) for lvl in dom.levels(root))
|
|
fitness = 1.0 / (1.0 + abs(12 - n_leaves))
|
|
return innerloop.Result(
|
|
x=np.array([0.25]), fitness=fitness, n_fails=0, fail_lines=(),
|
|
x0_fitness=fitness / 2, x0_n_fails=1, n_evals=budget, n_oracle_calls=1,
|
|
)
|
|
|
|
monkeypatch.setattr(innerloop, "optimise", fake_optimise)
|
|
init_root = dom.load(str(INIT_FILE))
|
|
r = driver.search(init_root, CORPUS, budget=4000, pop_size=4,
|
|
child_budget=60, seed=5, niche_by_signature=True,
|
|
restart_patience=300, restart_elite=1)
|
|
assert r.n_restarts >= 1
|
|
assert r.best is not None and r.best.fitness > 0
|
|
|
|
|
|
def test_feasibility_filter_off_matches_baseline(fake_inner):
|
|
"""§12.3: with the filter and reassociate OFF (defaults), the run is
|
|
identical to one that omits the params — a clean A/B control."""
|
|
init_root = dom.load(str(INIT_FILE))
|
|
base = driver.search(init_root, CORPUS, budget=600, pop_size=4,
|
|
child_budget=60, seed_budget=100, seed=9)
|
|
off = driver.search(init_root, CORPUS, budget=600, pop_size=4,
|
|
child_budget=60, seed_budget=100, seed=9,
|
|
enable_reassociate=False, feasibility_filter=False,
|
|
feasibility_max_shape_fails=0)
|
|
# Same search trajectory: identical best topology and accounting. (Absolute
|
|
# fitness carries the fake_inner monotone tiebreaker, which shares one call
|
|
# counter across both runs in this fixture, so compare the signature.)
|
|
assert off.best.sig == base.best.sig
|
|
assert off.n_topologies == base.n_topologies
|
|
assert off.n_evals == base.n_evals
|
|
|
|
|
|
def test_feasibility_filter_prunes_cheaply(fake_inner, monkeypatch):
|
|
"""§12.3 (homemaker-py-9gp.1): a pruned topology costs one feasibility eval
|
|
instead of the full child_budget, so the filter explores far more topologies
|
|
per budget; pruned individuals never displace the incumbent."""
|
|
from homemaker_layout import operators
|
|
|
|
# Force every filtered child to be pruned (shape-fail floor above any
|
|
# threshold and ≥ the incumbent's fail count).
|
|
monkeypatch.setattr(operators, "predicted_shape_fails",
|
|
lambda root, reqs, fit: 999)
|
|
|
|
init_root = dom.load(str(INIT_FILE))
|
|
budget, child_budget, pop_size = 1200, 60, 4
|
|
on = driver.search(init_root, CORPUS, budget=budget, pop_size=pop_size,
|
|
child_budget=child_budget, seed_budget=100, seed=4,
|
|
feasibility_filter=True, feasibility_max_shape_fails=0)
|
|
|
|
# Bootstrap (pop_size topologies at child_budget) then 1-eval prunes: the
|
|
# remaining budget buys ~one topology per eval, far more than child_budget.
|
|
bootstrap_evals = pop_size * child_budget
|
|
assert on.n_topologies > pop_size + (budget - bootstrap_evals) // child_budget
|
|
assert on.n_evals >= budget
|
|
# No pruned (untuned, fitness=0) individual is admitted to the population.
|
|
assert all(p.lineage and not p.lineage.startswith("pruned/") for p in on.population)
|
|
assert on.best is not None and not on.best.lineage.startswith("pruned/")
|
|
|
|
|
|
def test_search_parallel_smoke():
|
|
"""n_workers>1 runs without error and produces valid results."""
|
|
init_root = dom.load(str(INIT_FILE))
|
|
r = driver.search(init_root, CORPUS, budget=160, pop_size=2,
|
|
child_budget=80, seed=0, n_workers=2)
|
|
assert r.best is not None
|
|
assert r.best.fitness > 0
|
|
assert r.n_evals >= 160
|
|
assert 1 <= len(r.population) <= 2
|
|
assert r.n_topologies >= 2 # at least the bootstrap individuals
|
|
|
|
|
|
def test_search_parallel_is_reproducible():
|
|
"""Two same-seed parallel runs must be byte-identical (homemaker-py-xcy).
|
|
|
|
``_run_batch`` used to admit futures in completion order (``as_completed``),
|
|
which varies run-to-run; with the order-sensitive ``admit`` (n_evals accrual,
|
|
first-of-tie wins ``best``) that made parallel searches non-reproducible.
|
|
Admitting in submission order fixed it. Guard the invariant directly: same
|
|
seed + same worker count ⇒ identical best (n_fails, fitness, signature) and
|
|
identical improvement history."""
|
|
def run():
|
|
r = driver.search(dom.load(str(INIT_FILE)), CORPUS, budget=1200,
|
|
pop_size=8, child_budget=80, seed=0, n_workers=3)
|
|
return (r.best.n_fails, r.best.fitness, r.best.sig, tuple(r.history))
|
|
|
|
a = run()
|
|
b = run()
|
|
assert a == b, "parallel search is not reproducible run-to-run"
|
|
|
|
|
|
def _shared_best_result() -> driver.SearchResult:
|
|
"""A SearchResult whose best carries a live 3-room shared leaf (share=3),
|
|
plus a distinct C leaf — the harbor-house pathology in miniature."""
|
|
root = dom.Node(node=[[0, 0], [12, 0], [12, 8], [0, 8]],
|
|
height=2.7, wall_outer=0.25, wall_inner=0.08,
|
|
rotation=0, division=[0.5, 0.5])
|
|
root.left = dom.Node(type="n", share=3, share_type="n")
|
|
root.right = dom.Node(type="C")
|
|
dom._link(root)
|
|
best = driver.Individual(root=root, fitness=1e-5, n_fails=3, ratios={},
|
|
lineage="construct/0")
|
|
r = driver.SearchResult(best=best, population=[best], n_evals=1000,
|
|
n_topologies=5, n_distinct_signatures=4, n_restarts=1)
|
|
r.history = [(80, 1e-6, "construct/0"), (160, 1e-5, "core_divide noop")]
|
|
return r
|
|
|
|
|
|
def test_polish_finish_unfolds_and_stitches_rescore(fake_inner):
|
|
# homemaker-py-3l6, polish_budget<=0: unfold the shared leaf, rescore once
|
|
# under leaf_sharing off, and stitch accounting/history onto the sharing run.
|
|
r0 = _shared_best_result()
|
|
r = driver.polish_finish(r0, CORPUS, polish_budget=0, rescore_budget=150)
|
|
|
|
# the shared leaf is materialised into 3 distinct n rooms, stamps cleared
|
|
leaves = r.best.root.leaves()
|
|
assert sum(1 for lf in leaves if lf.type == "n") == 3
|
|
assert all(lf.share == 1 for lf in leaves)
|
|
# accounting is cumulative (1000 sharing evals + one 150-eval rescore)
|
|
assert r.n_evals == 1000 + 150
|
|
assert r.n_topologies == 5 + 1
|
|
assert r.n_distinct_signatures == 4 + 1
|
|
assert r.n_restarts == 1
|
|
# history keeps both phases, tagged so the objective change is visible
|
|
assert [lin for *_, lin in r.history[:2]] == [
|
|
"share:construct/0", "share:core_divide noop"]
|
|
assert r.history[-1][2].startswith("polish:")
|
|
# the rescore ran with leaf_sharing off (no sharing override reaches the inner)
|
|
assert not fake_inner[-1]["kw"].get("conf_overrides")
|
|
|
|
|
|
def test_polish_finish_runs_polish_search(fake_inner):
|
|
# polish_budget>0: a warm-started no-sharing search runs from the unfolded
|
|
# genome and its evals accrue on top of the sharing run.
|
|
r0 = _shared_best_result()
|
|
r = driver.polish_finish(r0, CORPUS, polish_budget=400, pop_size=3,
|
|
child_budget=80, seed=1)
|
|
assert r.n_evals > 1000 + 400 - 80 # sharing 1000 + ~400 polish evals
|
|
assert r.best.root.leaves() # a valid materialised genome survived
|
|
assert sum(1 for lf in r.best.root.leaves() if lf.share > 1) == 0
|
|
assert r.history[0][2].startswith("share:")
|
|
assert any(lin.startswith("polish:") for *_, lin in r.history)
|
|
|
|
|
|
def test_polish_finish_noop_without_best():
|
|
empty = driver.SearchResult(best=None, population=[], n_evals=0, n_topologies=0)
|
|
assert driver.polish_finish(empty, CORPUS, polish_budget=100) is empty
|
|
|
|
|
|
def test_search_seed_pop_evaluates_given_population(fake_inner):
|
|
# homemaker-py-kpu: seed_pop supplies an explicit initial population; each
|
|
# given root is evaluated (not bootstrapped/single-seeded) before the loop.
|
|
pop_roots = [dom.load(str(SEED_FILE)) for _ in range(3)]
|
|
r = driver.search(dom.load(str(INIT_FILE)), CORPUS, budget=0, pop_size=3,
|
|
child_budget=80, seed_budget=100, seed=0, seed_pop=pop_roots)
|
|
# budget 0 ⇒ only the 3 seed-pop evals run (100 each), no children
|
|
assert r.n_evals == 300
|
|
assert r.n_topologies == 3
|
|
assert all(ind.lineage.startswith("anneal-seed/") for ind in r.population)
|
|
|
|
|
|
def test_search_annealed_stitches_phases_and_finishes_honest(fake_inner):
|
|
# homemaker-py-kpu (Schedule B): the grain ramp runs one phase per ladder
|
|
# step plus a de-share polish, with cumulative accounting, a grain-tagged
|
|
# history, and a materialised (share-free) honest best.
|
|
r = driver.search_annealed(
|
|
dom.load(str(INIT_FILE)), CORPUS, budget=600, polish_budget=200,
|
|
grain_ladder=(3, 2), pop_size=3, child_budget=80, seed_budget=80, seed=0)
|
|
|
|
assert r.best is not None
|
|
# honest output: every shared leaf is materialised before the polish phase
|
|
assert all(lf.share == 1 for lf in r.best.root.leaves())
|
|
# accounting is cumulative across both sharing phases + polish
|
|
assert r.n_evals >= 600 + 200 - 80
|
|
# history is grain-tagged and ordered: first phase g3, then g2, then polish
|
|
tags = [lin.split(":", 1)[0] for *_, lin in r.history]
|
|
assert tags[0] == "g3"
|
|
assert "g2" in tags
|
|
assert tags[-1] == "polish"
|
|
# eval offsets are monotone non-decreasing across the stitched phases
|
|
evs = [e for e, *_ in r.history]
|
|
assert evs == sorted(evs)
|
|
|
|
|
|
def test_search_annealed_degenerate_ladder_falls_back(fake_inner):
|
|
# A ladder with no grain >= 2 has nothing to anneal: a plain no-sharing search
|
|
# over the full budget (+ polish), and the best is honest (share-free).
|
|
r = driver.search_annealed(
|
|
dom.load(str(INIT_FILE)), CORPUS, budget=300, polish_budget=100,
|
|
grain_ladder=(1,), pop_size=3, child_budget=80, seed_budget=80, seed=0)
|
|
assert r.best is not None
|
|
assert r.n_evals >= 300
|
|
assert all(lf.share == 1 for lf in r.best.root.leaves())
|