Spaces:
Sleeping
Sleeping
File size: 19,912 Bytes
d3955c7 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 | import os
import logging
import json
import numpy as np
import pandas as pd
from typing import List, Tuple, Dict
from dataclasses import dataclass, asdict
# Setup logging
logging.basicConfig(level=logging.INFO, format="%(asctime)s - %(levelname)s - %(message)s")
logger = logging.getLogger(__name__)
@dataclass
class GAParams:
"""Configurazione parametri GA."""
pop_size: int = 100
generations: int = 100
crossover_rate: float = 0.8
mutation_rate: float = 0.01
elitism_rate: float = 0.05
patience: int = 20
selection_method: str = "tournament"
crossover_method: str = "two-point"
mutation_method: str = "adaptive"
weights: Tuple[float, float, float] = (1.0, 0.4, 0.2)
tournament_size: int = 3
truncation_rate: float = 0.5
k_points: int = 2
max_freq_threshold: int = 16
seed: int = 42
load_gold_standard: bool = False # Mantenuto per compatibilità con l'API del Notebook
# Campi derivati (Properties) per evitare stati non sincronizzati
@property
def w_retention(self) -> float:
return self.weights[0]
@property
def w_penalty_freq(self) -> float:
return self.weights[1]
@property
def w_penalty_time(self) -> float:
return self.weights[2]
def to_dict(self):
return asdict(self)
@classmethod
def from_dict(cls, data):
# Mappatura sicura per ignorare chiavi extra o mappate diversamente
valid_keys = cls.__dataclass_fields__.keys()
filtered_data = {k: v for k, v in data.items() if k in valid_keys}
return cls(**filtered_data)
def save(self, path: str):
with open(path, 'w') as f:
json.dump(self.to_dict(), f, indent=4)
@classmethod
def load_gold_standard_config(cls, path: str = "ga_tuned_config.json") -> "GAParams":
"""Factory method pulito per caricare il JSON solo quando esplicitamente richiesto."""
if not os.path.exists(path):
logger.warning(f"File {path} non trovato. Fallback ai parametri di default.")
return cls()
try:
with open(path, 'r') as f:
config = json.load(f)
mapping = {
"selection_method": config.get("best_selection", "tournament"),
"crossover_method": config.get("best_crossover", "two-point"),
"mutation_method": config.get("best_mutation", "adaptive"),
"pop_size": config.get("opt_pop_size", 100),
"mutation_rate": config.get("opt_mutation_rate", 0.01),
"weights": tuple(config.get("weights", [1.0, 0.4, 0.2]))
}
return cls(**mapping)
except Exception as e:
logger.error(f"Errore caricamento JSON: {e}")
return cls()
class Chromosome:
"""Rappresenta una strategia di nudging (genotipo)."""
GENE_LENGTHS = {
'tipologia': 2,
'frequenza': 5,
'orario': 24
}
TOTAL_LENGTH = sum(GENE_LENGTHS.values())
def __init__(self, bits: np.ndarray = None, rng: np.random.Generator = None):
if bits is None:
# Blindiamo la riproducibilità usando il generatore RNG se fornito
if rng is not None:
self.bits = rng.integers(0, 2, self.TOTAL_LENGTH, dtype=np.int8)
else:
self.bits = np.random.randint(0, 2, self.TOTAL_LENGTH, dtype=np.int8)
else:
if len(bits) != self.TOTAL_LENGTH:
raise ValueError(f"Dimensione errata. Attesa: {self.TOTAL_LENGTH}, Trovata: {len(bits)}")
self.bits = np.array(bits, dtype=np.int8)
self._fitness = None
def decode(self) -> Dict:
"""Decodifica il genotipo nel fenotipo (parametri reali della strategia)."""
b = self.bits
tip_bits, freq_bits, ora_bits = b[0:2], b[2:7], b[7:31]
tipologia_idx = int(tip_bits.dot(1 << np.arange(tip_bits.size)[::-1]))
frequenza = int(freq_bits.dot(1 << np.arange(freq_bits.size)[::-1]))
tipologie = ["Promemoria", "Motivazionale", "Informativa", "Questionario"]
frequenza = max(1, frequenza) # Evita strategie da 0 notifiche
return {
'tipologia': tipologie[tipologia_idx],
'frequenza_settimanale': frequenza,
'orari_attivi': [h for h, val in enumerate(ora_bits) if val == 1],
'start_hour': next((h for h, val in enumerate(ora_bits) if val == 1), 9),
'end_hour': next((h for h, val in enumerate(reversed(ora_bits)) if val == 1), 18)
}
@property
def fitness(self):
return self._fitness
@fitness.setter
def fitness(self, value):
self._fitness = value
class FitnessEvaluator:
"""Calcola la fitness pesata a priori (Fase 0)."""
def __init__(self, patient_features: pd.Series, params: GAParams = None, rng: np.random.Generator = None):
self.patient_features = patient_features
self.params = params if params is not None else GAParams()
self.rng = rng if rng is not None else np.random.default_rng(self.params.seed)
# --- IL CUORE DEL BENCHMARK ---
# Questo contatore ci permette di fare un confronto "ad armi pari" con la Random Search
self.evaluation_calls = 0
def evaluate(self, chromosome: Chromosome, patient_features: pd.Series = None) -> float:
self.evaluation_calls += 1
fetch_features = patient_features if patient_features is not None else self.patient_features
phenotype = chromosome.decode()
retention_score = self._simulate_retention(phenotype, fetch_features)
mood_freq = fetch_features.get('mood_frequency_7d', 0.5)
dynamic_threshold = self.params.max_freq_threshold
if mood_freq > 0.8:
dynamic_threshold += 5
if mood_freq < 0.2:
dynamic_threshold -= 5
freq = phenotype['frequenza_settimanale']
# --- TASSA CONTINUA FREQUENZA ---
base_tax_f = (freq / 31.0) * 0.1 # Ogni messaggio ha un micro-costo di attenzione
penalty_freq = base_tax_f
if freq > dynamic_threshold:
diff = freq - dynamic_threshold
penalty_freq = min(1.0, base_tax_f + (np.exp(0.2 * diff) - 1) / 50) # Muro clinico
night_rate = fetch_features.get('night_activity_rate', 0.0)
night_hours = [23, 0, 1, 2, 3, 4, 5, 6]
active_hours = phenotype['orari_attivi']
# --- TASSA CONTINUA TEMPORALE ---
penalty_time = 0.0
if not active_hours:
penalty_time = 1.0
else:
# Ogni ora occupata ha una 'tassa di ingombro cognitivo'
base_tax_t = (len(active_hours) / 24.0) * 0.05
active_night_hours = sum(1 for h in active_hours if h in night_hours)
time_sensitivity = max(0.2, 1.0 - night_rate)
night_penalty = (active_night_hours / 8.0) * time_sensitivity
penalty_time = min(1.0, base_tax_t + night_penalty)
raw_fitness = (self.params.w_retention * retention_score) - \
(self.params.w_penalty_freq * penalty_freq) - \
(self.params.w_penalty_time * penalty_time)
return max(0.0001, float(raw_fitness))
def _simulate_retention(self, phenotype: Dict, features: pd.Series) -> float:
"""Simulatore del Patient Environment."""
mood_freq = features.get('mood_frequency_7d', 0.5)
avg_valence = features.get('avg_mood_valence_7d', 0.5)
read_rate = features.get('notification_read_rate', 0.5)
# 1. Identificazione Archetipi (Gerarchica e Coerente con Data Pipeline)
# Engaged: Il paziente ideale (Alta attività e umore stabile/positivo)
is_engaged = (mood_freq >= 0.6) and (avg_valence >= 0.5)
# Ghost: Il paziente che ha abbandonato (attività nulla o quasi)
is_ghost = not is_engaged and (mood_freq < 0.1) and (read_rate < 0.2)
# A Rischio: Il paziente in crisi (calo attività o umore negativo/preoccupante)
is_at_risk = not (is_engaged or is_ghost) and ((mood_freq < 0.3) or (avg_valence < 0.45))
# Moderato: Il paziente stabile, uso intermittente
is_moderato = not (is_engaged or is_ghost or is_at_risk)
score = 0.5
tipo = phenotype['tipologia']
freq = phenotype['frequenza_settimanale']
if is_engaged:
# Allineamento: L'engaged vuole mantenere l'abitudine (Promemoria)
if tipo == 'Promemoria':
score += 0.2
if 7 <= freq <= 14:
score += 0.2
if freq > 25:
score -= 0.3 # Anche l'engaged si stanca
elif is_at_risk:
# Allineamento: Chi è in crisi ha bisogno di motivazione o info
if tipo in ['Motivazionale', 'Informativa']:
score += 0.3
if 3 <= freq <= 7:
score += 0.2
if freq > 10:
score -= 0.2
elif is_ghost:
# Allineamento: Chi è sparito va recuperato con cautela
if tipo == 'Motivazionale':
score += 0.3
if freq <= 2:
score += 0.2
if freq > 5:
score -= 0.4 # Effetto spam garantito
elif is_moderato:
if tipo == 'Questionario':
score += 0.2
if tipo == 'Promemoria':
score += 0.1
if 4 <= freq <= 10:
score += 0.2
if freq > 25:
score -= 0.3
if len(phenotype['orari_attivi']) == 0:
score = 0.0
return max(0.0, min(1.0, score))
class GeneticAlgorithm:
"""Implementa il loop evolutivo con operatori configurabili (Fase 1 e 2)."""
def __init__(self, evaluator: FitnessEvaluator, params: GAParams, rng: np.random.Generator = None):
self.evaluator = evaluator
self.params = params
self.evaluator.params = self.params
self.population: List[Chromosome] = []
self.rng = rng if rng is not None else np.random.default_rng(self.params.seed)
self.history = {
"best_fitness": [],
"avg_fitness": [],
"diversity": []
}
self.current_diversity = 1.0
def initialize_population(self):
self.population = []
for _ in range(self.params.pop_size):
# Passiamo l'RNG locale per evitare leakage stocastico
self.population.append(Chromosome(rng=self.rng))
self._evaluate_population()
def _evaluate_population(self):
# Valuta solo chi non ha la fitness (Risparmio computazionale)
for ind in self.population:
if ind.fitness is None:
ind.fitness = self.evaluator.evaluate(ind)
self.population.sort(key=lambda x: x.fitness, reverse=True)
# --- SELECTION METHODS ---
def _select(self) -> Chromosome:
method = self.params.selection_method.lower()
if method == "tournament":
return self._selection_tournament()
elif method == "roulette":
return self._selection_roulette()
elif method == "ranking":
return self._selection_ranking()
elif method == "truncation":
return self._selection_truncation()
else:
return self._selection_roulette()
def _selection_tournament(self) -> Chromosome:
candidates = self.rng.choice(self.population, size=self.params.tournament_size, replace=False)
best = max(candidates, key=lambda x: x.fitness)
new_ind = Chromosome(bits=best.bits.copy())
new_ind.fitness = best.fitness
return new_ind
def _selection_roulette(self) -> Chromosome:
fitnesses = np.array([max(0, ind.fitness) for ind in self.population])
total = sum(fitnesses)
if total == 0:
idx = self.rng.integers(0, len(self.population))
new_ind = Chromosome(bits=self.population[idx].bits.copy())
new_ind.fitness = self.population[idx].fitness
return new_ind
probs = fitnesses / total
idx = self.rng.choice(len(self.population), p=probs)
new_ind = Chromosome(bits=self.population[idx].bits.copy())
new_ind.fitness = self.population[idx].fitness
return new_ind
def _selection_truncation(self) -> Chromosome:
cutoff = max(1, int(len(self.population) * self.params.truncation_rate))
best_set = self.population[:cutoff]
idx = self.rng.integers(0, len(best_set))
new_ind = Chromosome(bits=best_set[idx].bits.copy())
new_ind.fitness = best_set[idx].fitness
return new_ind
def _selection_ranking(self) -> Chromosome:
n = len(self.population)
ranks = np.arange(n, 0, -1)
total_ranks = sum(ranks)
probs = ranks / total_ranks
idx = self.rng.choice(n, p=probs)
new_ind = Chromosome(bits=self.population[idx].bits.copy())
new_ind.fitness = self.population[idx].fitness
return new_ind
# --- CROSSOVER METHODS ---
def _crossover(self, p1: Chromosome, p2: Chromosome) -> Tuple[Chromosome, Chromosome]:
if self.rng.random() > self.params.crossover_rate:
# RISPARMIO COMPTUAZIONALE: Se non c'è crossover, passiamo la fitness in eredità intatta
c1, c2 = Chromosome(bits=p1.bits.copy()), Chromosome(bits=p2.bits.copy())
c1.fitness, c2.fitness = p1.fitness, p2.fitness
return c1, c2
method = self.params.crossover_method.lower()
if method == "single-point":
return self._crossover_1point(p1, p2)
elif method == "two-point":
return self._crossover_2point(p1, p2)
elif method == "uniform":
return self._crossover_uniform(p1, p2)
elif method == "k-point":
return self._crossover_kpoint(p1, p2)
else:
return self._crossover_1point(p1, p2)
def _crossover_1point(self, p1: Chromosome, p2: Chromosome) -> Tuple[Chromosome, Chromosome]:
pt = self.rng.integers(1, Chromosome.TOTAL_LENGTH)
c1 = np.concatenate((p1.bits[:pt], p2.bits[pt:]))
c2 = np.concatenate((p2.bits[:pt], p1.bits[pt:]))
return Chromosome(bits=c1), Chromosome(bits=c2)
def _crossover_2point(self, p1: Chromosome, p2: Chromosome) -> Tuple[Chromosome, Chromosome]:
pt1 = self.rng.integers(1, Chromosome.TOTAL_LENGTH - 1)
pt2 = self.rng.integers(pt1 + 1, Chromosome.TOTAL_LENGTH)
c1 = np.concatenate((p1.bits[:pt1], p2.bits[pt1:pt2], p1.bits[pt2:]))
c2 = np.concatenate((p2.bits[:pt1], p1.bits[pt1:pt2], p2.bits[pt2:]))
return Chromosome(bits=c1), Chromosome(bits=c2)
def _crossover_uniform(self, p1: Chromosome, p2: Chromosome) -> Tuple[Chromosome, Chromosome]:
mask = self.rng.integers(0, 2, Chromosome.TOTAL_LENGTH)
c1 = np.where(mask == 1, p1.bits, p2.bits)
c2 = np.where(mask == 1, p2.bits, p1.bits)
return Chromosome(bits=c1), Chromosome(bits=c2)
def _crossover_kpoint(self, p1: Chromosome, p2: Chromosome) -> Tuple[Chromosome, Chromosome]:
pts = sorted(self.rng.choice(range(1, Chromosome.TOTAL_LENGTH), size=self.params.k_points, replace=False))
pts = [0] + list(pts) + [Chromosome.TOTAL_LENGTH]
c1_bits, c2_bits = [], []
for i in range(len(pts)-1):
if i % 2 == 0:
c1_bits.append(p1.bits[pts[i]:pts[i+1]])
c2_bits.append(p2.bits[pts[i]:pts[i+1]])
else:
c1_bits.append(p2.bits[pts[i]:pts[i+1]])
c2_bits.append(p1.bits[pts[i]:pts[i+1]])
return Chromosome(bits=np.concatenate(c1_bits)), Chromosome(bits=np.concatenate(c2_bits))
# --- MUTATION METHODS ---
def _mutate(self, ind: Chromosome) -> None:
"""Esegue la mutazione. Azzera la fitness SOLO se c'è stato un reale cambiamento del DNA."""
method = self.params.mutation_method.lower()
mutated = False
if method == "flip-bit":
mutated = self._mutation_flip(ind)
elif method == "multi-bit":
mutated = self._mutation_multi(ind)
elif method == "adaptive":
mutated = self._mutation_adaptive(ind)
else:
mutated = self._mutation_flip(ind)
if mutated:
ind.fitness = None # Invalida la cache
def _mutation_flip(self, ind: Chromosome) -> bool:
mutated = False
for i in range(Chromosome.TOTAL_LENGTH):
if self.rng.random() < self.params.mutation_rate:
ind.bits[i] = 1 - ind.bits[i]
mutated = True
return mutated
def _mutation_multi(self, ind: Chromosome) -> bool:
k = self.rng.integers(1, 4, endpoint=True)
indices = self.rng.choice(range(Chromosome.TOTAL_LENGTH), size=k, replace=False)
for idx in indices:
ind.bits[idx] = 1 - ind.bits[idx]
return True
def _mutation_adaptive(self, ind: Chromosome) -> bool:
adj_rate = self.params.mutation_rate
if self.current_diversity < 0.1:
adj_rate *= 2.0
elif self.current_diversity > 0.4:
adj_rate *= 0.5
mutated = False
for i in range(Chromosome.TOTAL_LENGTH):
if self.rng.random() < adj_rate:
ind.bits[i] = 1 - ind.bits[i]
mutated = True
return mutated
def _calculate_diversity(self) -> float:
pop_matrix = np.array([ind.bits for ind in self.population])
p1 = pop_matrix.mean(axis=0)
avg_hamming = np.sum(2 * p1 * (1 - p1))
return float(avg_hamming / Chromosome.TOTAL_LENGTH)
def run(self):
self.initialize_population()
n_elites = max(1, int(self.params.pop_size * self.params.elitism_rate))
for gen in range(self.params.generations):
self.current_diversity = self._calculate_diversity()
new_population = []
# Elitarismo: i migliori passano incondizionatamente con la fitness già calcolata!
for i in range(n_elites):
elite_ind = Chromosome(bits=self.population[i].bits.copy())
elite_ind.fitness = self.population[i].fitness
new_population.append(elite_ind)
# Riproduzione
while len(new_population) < self.params.pop_size:
p1 = self._select()
p2 = self._select()
c1, c2 = self._crossover(p1, p2)
self._mutate(c1)
self._mutate(c2)
new_population.extend([c1, c2])
self.population = new_population[:self.params.pop_size]
self._evaluate_population()
# Statistiche
best_fit = self.population[0].fitness
avg_fit = sum(ind.fitness for ind in self.population) / self.params.pop_size
self.history["best_fitness"].append(best_fit)
self.history["avg_fitness"].append(avg_fit)
self.history["diversity"].append(self.current_diversity)
# Early Stopping (Fase 1.5)
if gen >= self.params.patience:
recent_bests = self.history["best_fitness"][-self.params.patience:]
if (recent_bests[-1] - recent_bests[0]) < 1e-6:
logger.debug(f"Early Stopping alla generazione {gen} per mancanza di miglioramento.")
break
return self.population[0] |