File size: 6,287 Bytes
c289d87 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 | from __future__ import annotations
from dataclasses import dataclass
from typing import Dict, Iterable, List, Sequence
import numpy as np
@dataclass
class EpsilonRegretConfig:
epsilon_quality: float = 0.10
acceptable_regret: float = 0.10
min_budget: int = 250
max_budget: int = 10000
min_model_weight: float = 0.25
patience_rounds: int = 6
min_improvement: float = 0.02
uncertainty_multiplier: float = 0.50
min_cluster_coverage: float = 0.20
min_hypercluster_coverage: float = 0.20
class EpsilonRegretController:
"""
Budget-aware stop controller with epsilon-regret semantics.
Online intent:
stop when estimated remaining gain is small relative to current best score,
after minimum exploration and coverage constraints are satisfied.
"""
def __init__(self, config: EpsilonRegretConfig | None = None) -> None:
self.config = config or EpsilonRegretConfig()
self.best_score = np.inf
self.stagnant_rounds = 0
self.history: List[Dict[str, float | int | bool | str]] = []
@staticmethod
def _mean_or_nan(values: Sequence[float]) -> float:
arr = np.asarray(values, dtype=float)
if arr.size == 0:
return float("nan")
return float(np.nanmean(arr))
@staticmethod
def _estimate_remaining_gain(
best_score: float,
expected_remaining_scores: Iterable[float],
uncertainty_scores: Iterable[float],
uncertainty_multiplier: float,
) -> float:
pred = np.asarray(list(expected_remaining_scores), dtype=float)
unc = np.asarray(list(uncertainty_scores), dtype=float)
if pred.size == 0:
return float("inf")
if unc.size != pred.size:
unc = np.full(pred.shape[0], np.nanmean(unc) if unc.size else 1.0, dtype=float)
optimistic = pred - float(max(0.0, uncertainty_multiplier)) * np.abs(unc)
optimistic_best = float(np.nanmin(optimistic)) if np.isfinite(optimistic).any() else float(np.nanmin(pred))
if not np.isfinite(optimistic_best) or not np.isfinite(best_score):
return float("inf")
# lower score is better
return float(max(0.0, best_score - optimistic_best))
def update(
self,
*,
round_idx: int,
n_evaluated: int,
best_score_now: float,
recent_scores: Sequence[float],
model_weight: float,
cluster_coverage: float,
hypercluster_coverage: float,
expected_remaining_scores: Iterable[float],
uncertainty_scores: Iterable[float],
time_importance: float,
budget_cap: int,
) -> Dict[str, float | int | bool | str]:
cfg = self.config
ti = float(max(0.0, min(1.0, float(time_importance))))
if np.isfinite(self.best_score):
improvement = float(self.best_score - best_score_now)
else:
improvement = float("inf")
if improvement > float(cfg.min_improvement):
self.stagnant_rounds = 0
else:
self.stagnant_rounds += 1
self.best_score = min(float(self.best_score), float(best_score_now))
epsilon_rel = float(max(cfg.epsilon_quality, cfg.acceptable_regret))
# Higher time importance permits slightly larger acceptable residual gain.
epsilon_rel *= float(1.0 + 0.5 * ti)
epsilon_abs = float(epsilon_rel * max(1.0, abs(float(self.best_score))))
remaining_gain = self._estimate_remaining_gain(
best_score=float(self.best_score),
expected_remaining_scores=expected_remaining_scores,
uncertainty_scores=uncertainty_scores,
uncertainty_multiplier=float(cfg.uncertainty_multiplier),
)
regret_ratio_proxy = float(remaining_gain / max(1.0, abs(float(self.best_score))))
budget_floor = int(max(1, cfg.min_budget))
budget_ceiling = int(min(cfg.max_budget, budget_cap))
budget_ready = int(n_evaluated) >= budget_floor
model_ready = float(model_weight) >= float(cfg.min_model_weight)
coverage_ready = bool(
float(cluster_coverage) >= float(cfg.min_cluster_coverage)
and float(hypercluster_coverage) >= float(cfg.min_hypercluster_coverage)
)
stagnation_ready = int(self.stagnant_rounds) >= int(cfg.patience_rounds)
hit_budget_ceiling = int(n_evaluated) >= int(budget_ceiling)
epsilon_ready = bool(np.isfinite(remaining_gain) and remaining_gain <= epsilon_abs)
stop = bool(hit_budget_ceiling or (budget_ready and model_ready and coverage_ready and stagnation_ready and epsilon_ready))
if hit_budget_ceiling:
reason = "max_budget_reached"
elif not budget_ready:
reason = "below_min_budget"
elif not model_ready:
reason = "model_not_ready"
elif not coverage_ready:
reason = "coverage_not_ready"
elif not stagnation_ready:
reason = "insufficient_stagnation_evidence"
elif not epsilon_ready:
reason = "remaining_gain_above_epsilon"
else:
reason = "epsilon_regret_stop"
rec = {
"round_idx": int(round_idx),
"n_evaluated": int(n_evaluated),
"best_score": float(self.best_score),
"recent_mean_score": self._mean_or_nan(recent_scores),
"model_weight": float(model_weight),
"cluster_coverage": float(cluster_coverage),
"hypercluster_coverage": float(hypercluster_coverage),
"stagnant_rounds": int(self.stagnant_rounds),
"epsilon_abs": float(epsilon_abs),
"estimated_remaining_gain": float(remaining_gain),
"regret_ratio_proxy": float(regret_ratio_proxy),
"budget_floor": int(budget_floor),
"budget_ceiling": int(budget_ceiling),
"budget_ready": bool(budget_ready),
"model_ready": bool(model_ready),
"coverage_ready": bool(coverage_ready),
"stagnation_ready": bool(stagnation_ready),
"epsilon_ready": bool(epsilon_ready),
"stop": bool(stop),
"reason": reason,
}
self.history.append(rec)
return rec
|