Bruno Postle
d08d15e4d7
Driving equal-offset cut ratios with Nelder-Mead against the REAL oracle fitness (full objective, no proxy) improves all three test candidates with zero new failures: 2f45907 (best evolved) 0.012617 -> 0.015684 x1.24 (2->2 fails) candidate-002 0.007375 -> 0.012319 x1.67 (2->2 fails) c964435 (baseline) 0.003667 -> 0.005836 x1.59 (3->3 fails) Headroom widens on weaker designs. The EA under-optimises geometry by 24-67% even on its best result. This validates a full-fitness geometry inner loop (NOT the earlier area-proxy solver) and motivates a memetic architecture: topology search outside, full-objective geometry optimise inside, gated on a native Python fitness (oracle at ~3s/call is too slow).
89 lines
2.8 KiB
Python
89 lines
2.8 KiB
Python
"""Full-fitness frozen-topology optimisation.
|
|
|
|
Drive the equal-offset division ratios with a derivative-free optimiser against
|
|
the REAL oracle fitness (the whole objective, not an area proxy) on a fixed
|
|
topology. This removes both confounds of the earlier sweeps (partial objective,
|
|
proxy target) and answers the only question that matters next:
|
|
|
|
Is there geometry headroom above the EA's designs, or are they already
|
|
geometry-optima (=> the bottleneck is topology + the 0.5^n cliff)?
|
|
|
|
For each candidate: report fitness/fails before and after.
|
|
"""
|
|
|
|
import sys
|
|
from pathlib import Path
|
|
|
|
import numpy as np
|
|
from scipy.optimize import minimize
|
|
|
|
sys.path.insert(0, str(Path(__file__).resolve().parents[1] / "src"))
|
|
|
|
from homemaker import dom, oracle, solver # noqa: E402
|
|
|
|
URB = Path("/home/bruno/src/urb")
|
|
EX = URB / "examples/programme-house"
|
|
SCRATCH = Path(__file__).resolve().parents[1] / "scratch"
|
|
_EPS = 0.02
|
|
|
|
CANDIDATES = [
|
|
"candidate-002.dom", # ~0.0074 (MCP-refined)
|
|
"c964435454c459f86c3ed9a5a7621132.dom", # ~0.0037 (MCP baseline)
|
|
]
|
|
|
|
|
|
def optimise(src: Path, maxfev: int = 200):
|
|
import shutil
|
|
|
|
shutil.copy(src, SCRATCH / "orig.dom")
|
|
s0 = oracle.score(SCRATCH / "orig.dom", URB)
|
|
|
|
root = dom.load(str(src))
|
|
free = solver.free_branches(root)
|
|
x0 = np.array([b.division[0] for b in free], dtype=float)
|
|
opt_path = SCRATCH / "opt.dom"
|
|
|
|
best = {"f": s0.fitness, "fails": s0.n_fails, "x": x0.copy()}
|
|
|
|
def neg_fitness(x: np.ndarray) -> float:
|
|
xc = np.clip(x, _EPS, 1 - _EPS)
|
|
for j, b in enumerate(free):
|
|
b.division = [float(xc[j]), float(xc[j])]
|
|
dom.dump(root, str(opt_path))
|
|
try:
|
|
s = oracle.score(opt_path, URB)
|
|
except Exception: # noqa: BLE001
|
|
return 1e9
|
|
if s.fitness > best["f"]:
|
|
best.update(f=s.fitness, fails=s.n_fails, x=xc.copy())
|
|
return -s.fitness
|
|
|
|
minimize(
|
|
neg_fitness, x0, method="Nelder-Mead",
|
|
options={"maxfev": maxfev, "xatol": 1e-3, "fatol": 1e-12},
|
|
)
|
|
return s0, best, len(free)
|
|
|
|
|
|
def main() -> None:
|
|
SCRATCH.mkdir(exist_ok=True)
|
|
import shutil
|
|
|
|
shutil.copy(EX / "patterns.config", SCRATCH / "patterns.config")
|
|
|
|
maxfev = int(sys.argv[1]) if len(sys.argv) > 1 else 200
|
|
for name in CANDIDATES:
|
|
s0, best, ndof = optimise(EX / name, maxfev)
|
|
ratio = best["f"] / s0.fitness if s0.fitness else float("inf")
|
|
verdict = "IMPROVED" if best["f"] > s0.fitness * 1.001 else "no gain"
|
|
print(
|
|
f"{name:42s} dof={ndof:2d} "
|
|
f"orig={s0.fitness:.5g}(fails {s0.n_fails}) "
|
|
f"opt={best['f']:.5g}(fails {best['fails']}) "
|
|
f"x{ratio:.2f} {verdict}",
|
|
flush=True,
|
|
)
|
|
|
|
|
|
if __name__ == "__main__":
|
|
main()
|