""" PHOENIX quantum feasibility sweep -- shared harness. Purpose (the honest mandate): screen EVERY simulable quantum / quantum-inspired method on the untested list against a MANDATORY classical baseline, on real-shaped PHOENIX fire-detection artifacts, and test whether STACKING several of them creates any new value a single method (or the classical baseline) does not already provide. Discipline (same as the research log): - every quantum method is scored head-to-head with a classical baseline on the same data by the same metric; - a shuffled-label PLACEBO accompanies every learned-separability screen; - nothing is promoted; "no advantage" is a valid, reported result; - feasibility screens run on a realistic SEVIRI-calibrated dataset (derived features ~ chance, full 11-band spectrum ~ 0.63 AUC, matching entry 0088); anything that beats classical here escalates to the REAL spec12 corpus on RunPod before any claim. Compute: CPU only, small (n<=~120, <=8 qubits), laptop-light, load-aware. No QPU, no DGX. Real-QPU is reached only if a screen survives the simulator. """ from __future__ import annotations import time import numpy as np from dataclasses import dataclass, field from sklearn.svm import SVC from sklearn.ensemble import GradientBoostingClassifier from sklearn.linear_model import LogisticRegression from sklearn.model_selection import StratifiedKFold, cross_val_predict from sklearn.preprocessing import StandardScaler from sklearn.metrics import roc_auc_score BANDS = ["ir039", "wv062", "wv073", "ir087", "ir097", "ir108", "ir120", "ir134", "vis006", "vis008", "nir016"] # -------------------------------------------------------------------------- # Realistic SEVIRI-calibrated fire / non-fire dataset # -------------------------------------------------------------------------- def make_seviri_dataset(n: int = 120, seed: int = 0, signal: float = 0.85): """ 11-band fire vs intensity-matched non-fire pixels. Construction goal (calibrated to entry 0088): the three DERIVED features (temperature ~ ir039, MIR-TIR ~ ir039-ir108, spatial-z proxy) do NOT separate (intensity-matched), but a genuine combustion direction spread across the IR bands gives a modest full-spectrum signal. Non-fire distractors (hot ground, sun-glint) mimic the derived features without the full spectral structure. """ rng = np.random.default_rng(seed) half = n // 2 # shared thermal background (intensity), identical distribution for both classes base_temp = rng.normal(295, 6, n) # K-ish X = np.zeros((n, 11)) X[:, 0] = base_temp + rng.normal(0, 2, n) # ir039 (MIR) driven by temp X[:, 5] = base_temp + rng.normal(0, 2, n) # ir108 (TIR) driven by temp for b in [1, 2, 3, 4, 6, 7]: # other IR bands X[:, b] = base_temp * rng.uniform(0.9, 1.0) + rng.normal(0, 3, n) for b in [8, 9, 10]: # visible/NIR: solar, class-blind X[:, b] = rng.normal(40, 12, n) y = np.zeros(n, int); y[:half] = 1 rng.shuffle(y) # genuine combustion direction: kept BLIND to the derived features by giving ir039 # (idx0) and ir108 (idx5) equal, small weight (so temperature and MIR-TIR net ~0), # and placing the real signal in the other IR/WV bands -> only the full spectrum reads it comb_dir = np.array([0.15, 0.30, 0.30, 0.50, 0.25, 0.15, 0.45, 0.40, 0.0, 0.0, 0.0]) comb_dir = comb_dir / np.linalg.norm(comb_dir) X += signal * np.outer(y, comb_dir) * 6.0 # intensity-matched distractors: a CLASS-BLIND subset (both fire and non-fire) get a # "hot" spike on ir039 & MIR-TIR -> intensity/temperature cannot read the label distract = rng.random(n) < 0.35 X[distract, 0] += rng.uniform(3, 8, distract.sum()) X[distract, 5] -= rng.uniform(0, 3, distract.sum()) X += rng.normal(0, 1.5, X.shape) return X, y def derived_features(X): """The three summary features PHOENIX historically scored on (entry 0088).""" temp = X[:, 0] mir_tir = X[:, 0] - X[:, 5] spatial_z = (X[:, 0] - X[:, 0].mean()) / (X[:, 0].std() + 1e-9) return np.stack([temp, mir_tir, spatial_z], axis=1) # -------------------------------------------------------------------------- # Scoring helpers # -------------------------------------------------------------------------- def cv_auc(model, X, y, seed: int = 0, folds: int = 5) -> float: skf = StratifiedKFold(n_splits=folds, shuffle=True, random_state=seed) try: if hasattr(model, "decision_function"): s = cross_val_predict(model, X, y, cv=skf, method="decision_function") else: s = cross_val_predict(model, X, y, cv=skf, method="predict_proba")[:, 1] except Exception: s = cross_val_predict(model, X, y, cv=skf, method="predict_proba")[:, 1] return float(roc_auc_score(y, s)) def auc_from_scores(y, scores) -> float: return float(roc_auc_score(y, scores)) def placebo_auc(model_fn, X, y, seed: int = 0) -> float: rng = np.random.default_rng(seed + 999) yp = y.copy(); rng.shuffle(yp) return cv_auc(model_fn(), X, yp, seed=seed) @dataclass class ScreenResult: name: str # method (list item) task: str quantum_metric: float classical_metric: float placebo: float | None metric_name: str latency_s: float verdict: str = "" extra: dict = field(default_factory=dict) def advantage(self, tol: float = 0.01) -> bool: return self.quantum_metric > self.classical_metric + tol def classical_baselines(X, y, seed=0) -> dict: """Mandatory classical baselines: full-spectrum and derived-only.""" Xs = StandardScaler().fit_transform(X) out = { "rbf_svm_full": cv_auc(SVC(kernel="rbf", C=2.0), Xs, y, seed), "gbm_full": cv_auc(GradientBoostingClassifier(random_state=seed), Xs, y, seed), "logreg_full": cv_auc(LogisticRegression(max_iter=500), Xs, y, seed), "rbf_svm_derived": cv_auc(SVC(kernel="rbf", C=2.0), StandardScaler().fit_transform(derived_features(X)), y, seed), } out["best_classical_full"] = max(out["rbf_svm_full"], out["gbm_full"], out["logreg_full"]) return out def verdict_line(r: ScreenResult) -> str: if r.advantage(): return (f"POSSIBLE lift: {r.metric_name} {r.quantum_metric:.3f} vs classical " f"{r.classical_metric:.3f} (+{r.quantum_metric - r.classical_metric:.3f}) " f"-> escalate to real corpus before any claim.") return (f"No advantage: {r.metric_name} {r.quantum_metric:.3f} vs classical " f"{r.classical_metric:.3f}. Classical baseline holds.")