Upload mle/inference.py
Browse files- mle/inference.py +446 -0
mle/inference.py
ADDED
|
@@ -0,0 +1,446 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""
|
| 2 |
+
Moteur d'Inférence avec Apprentissage en Ligne
|
| 3 |
+
|
| 4 |
+
L'inférence minimise l'énergie par descente stochastique locale.
|
| 5 |
+
À chaque itération :
|
| 6 |
+
1. Calcule les voisins via le routeur
|
| 7 |
+
2. Évalue l'énergie du paysage
|
| 8 |
+
3. Sélectionne les flips de bits qui réduisent l'énergie
|
| 9 |
+
4. Met à jour les associations (apprentissage en ligne)
|
| 10 |
+
5. Détecte motifs pour abstraction
|
| 11 |
+
|
| 12 |
+
La minimisation est un processus de Monte Carlo / Hopfield-like
|
| 13 |
+
mais avec mémoire adaptative et apprentissage continu.
|
| 14 |
+
"""
|
| 15 |
+
|
| 16 |
+
import numpy as np
|
| 17 |
+
from numba import njit, prange
|
| 18 |
+
from typing import List, Tuple, Dict, Optional, Callable
|
| 19 |
+
import logging
|
| 20 |
+
import time
|
| 21 |
+
|
| 22 |
+
logger = logging.getLogger(__name__)
|
| 23 |
+
|
| 24 |
+
VECTOR_SIZE = 4096
|
| 25 |
+
|
| 26 |
+
|
| 27 |
+
@njit(cache=True)
|
| 28 |
+
def random_flip_batch(state: np.ndarray, n_flips: int, rng_seed: int) -> np.ndarray:
|
| 29 |
+
"""Flip aléatoire de n_flips bits."""
|
| 30 |
+
np.random.seed(rng_seed)
|
| 31 |
+
new_state = state.copy()
|
| 32 |
+
flip_indices = np.random.choice(VECTOR_SIZE, size=n_flips, replace=False)
|
| 33 |
+
for idx in flip_indices:
|
| 34 |
+
new_state[idx] = 1 - new_state[idx]
|
| 35 |
+
return new_state
|
| 36 |
+
|
| 37 |
+
|
| 38 |
+
@njit(cache=True)
|
| 39 |
+
def hamming_distance(a: np.ndarray, b: np.ndarray) -> int:
|
| 40 |
+
"""Distance de Hamming entre deux vecteurs binaires."""
|
| 41 |
+
dist = 0
|
| 42 |
+
for i in range(len(a)):
|
| 43 |
+
dist += a[i] ^ b[i]
|
| 44 |
+
return dist
|
| 45 |
+
|
| 46 |
+
|
| 47 |
+
class InferenceResult:
|
| 48 |
+
"""Résultat d'une inférence complète."""
|
| 49 |
+
|
| 50 |
+
def __init__(self):
|
| 51 |
+
self.initial_state: Optional[np.ndarray] = None
|
| 52 |
+
self.final_state: Optional[np.ndarray] = None
|
| 53 |
+
self.energy_trajectory: List[float] = []
|
| 54 |
+
self.neighbor_trajectory: List[List[Tuple[int, float]]] = []
|
| 55 |
+
self.n_iterations = 0
|
| 56 |
+
self.converged = False
|
| 57 |
+
self.creation_events: List[Dict] = []
|
| 58 |
+
self.learning_events: List[Dict] = []
|
| 59 |
+
self.execution_time_ms = 0.0
|
| 60 |
+
|
| 61 |
+
|
| 62 |
+
class InferenceEngine:
|
| 63 |
+
"""
|
| 64 |
+
Moteur d'inférence par minimisation d'énergie avec apprentissage en ligne.
|
| 65 |
+
|
| 66 |
+
Paramètres clés:
|
| 67 |
+
- temperature: contrôle le bruit dans la descente (plus haut = plus exploratoire)
|
| 68 |
+
- max_iterations: nombre max d'itérations de minimisation
|
| 69 |
+
- energy_tolerance: seuil de convergence
|
| 70 |
+
- learning_rate: vitesse d'apprentissage pendant l'inférence
|
| 71 |
+
"""
|
| 72 |
+
|
| 73 |
+
def __init__(
|
| 74 |
+
self,
|
| 75 |
+
temperature: float = 0.5,
|
| 76 |
+
cooling_rate: float = 0.995,
|
| 77 |
+
max_iterations: int = 100,
|
| 78 |
+
energy_tolerance: float = 1.0,
|
| 79 |
+
learning_rate: float = 0.01,
|
| 80 |
+
online_learning: bool = True,
|
| 81 |
+
pattern_detection_interval: int = 10,
|
| 82 |
+
convergence_window: int = 5,
|
| 83 |
+
early_stop_threshold: float = 0.001,
|
| 84 |
+
):
|
| 85 |
+
self.temperature = temperature
|
| 86 |
+
self.cooling_rate = cooling_rate
|
| 87 |
+
self.max_iterations = max_iterations
|
| 88 |
+
self.energy_tolerance = energy_tolerance
|
| 89 |
+
self.learning_rate = learning_rate
|
| 90 |
+
self.online_learning = online_learning
|
| 91 |
+
self.pattern_detection_interval = pattern_detection_interval
|
| 92 |
+
self.convergence_window = convergence_window
|
| 93 |
+
self.early_stop_threshold = early_stop_threshold
|
| 94 |
+
|
| 95 |
+
# Stats
|
| 96 |
+
self.total_inferences = 0
|
| 97 |
+
self.total_iterations = 0
|
| 98 |
+
self.total_converged = 0
|
| 99 |
+
self.avg_inference_time_ms = 0.0
|
| 100 |
+
|
| 101 |
+
def infer(
|
| 102 |
+
self,
|
| 103 |
+
initial_state: np.ndarray,
|
| 104 |
+
memory_table,
|
| 105 |
+
router,
|
| 106 |
+
energy_landscape,
|
| 107 |
+
binder,
|
| 108 |
+
k_neighbors: int = 10,
|
| 109 |
+
external_callback: Optional[Callable] = None,
|
| 110 |
+
) -> InferenceResult:
|
| 111 |
+
"""
|
| 112 |
+
Inférence complète avec minimisation d'énergie et apprentissage en ligne.
|
| 113 |
+
|
| 114 |
+
Args:
|
| 115 |
+
initial_state: état initial du système (4096 bits)
|
| 116 |
+
memory_table: SparseAddressTable
|
| 117 |
+
router: HammingRouter
|
| 118 |
+
energy_landscape: EnergyLandscape
|
| 119 |
+
binder: CircularBinder
|
| 120 |
+
k_neighbors: nombre de voisins à considérer
|
| 121 |
+
external_callback: fonction optionnelle appelée à chaque itération
|
| 122 |
+
|
| 123 |
+
Returns:
|
| 124 |
+
InferenceResult avec trajectoire et événements d'apprentissage
|
| 125 |
+
"""
|
| 126 |
+
import time
|
| 127 |
+
t0 = time.time()
|
| 128 |
+
|
| 129 |
+
result = InferenceResult()
|
| 130 |
+
result.initial_state = initial_state.copy()
|
| 131 |
+
|
| 132 |
+
current_state = initial_state.copy()
|
| 133 |
+
temperature = self.temperature
|
| 134 |
+
|
| 135 |
+
prev_energy = float('inf')
|
| 136 |
+
energy_window = []
|
| 137 |
+
|
| 138 |
+
# Trajectoire des états pour détection de motifs
|
| 139 |
+
state_trajectory = [current_state.copy()]
|
| 140 |
+
|
| 141 |
+
for iteration in range(self.max_iterations):
|
| 142 |
+
# 1. Route vers les voisins les plus proches
|
| 143 |
+
neighbors_info = router.route(
|
| 144 |
+
current_state,
|
| 145 |
+
k=k_neighbors,
|
| 146 |
+
use_cache=True
|
| 147 |
+
)
|
| 148 |
+
|
| 149 |
+
if len(neighbors_info) == 0:
|
| 150 |
+
# Pas de voisins : état nouveau, potentiellement créer
|
| 151 |
+
break
|
| 152 |
+
|
| 153 |
+
neighbor_indices = [idx for idx, _ in neighbors_info]
|
| 154 |
+
neighbor_distances = [dist for _, dist in neighbors_info]
|
| 155 |
+
|
| 156 |
+
# Récupère les vecteurs voisins depuis la mémoire
|
| 157 |
+
neighbor_vectors = np.array([
|
| 158 |
+
memory_table.vectors[idx]
|
| 159 |
+
for idx in neighbor_indices
|
| 160 |
+
if memory_table.active_mask[idx]
|
| 161 |
+
], dtype=np.uint8)
|
| 162 |
+
|
| 163 |
+
if len(neighbor_vectors) == 0:
|
| 164 |
+
break
|
| 165 |
+
|
| 166 |
+
neighbor_ids = [
|
| 167 |
+
memory_table.metadata[idx].id
|
| 168 |
+
for idx in neighbor_indices
|
| 169 |
+
if memory_table.active_mask[idx]
|
| 170 |
+
]
|
| 171 |
+
|
| 172 |
+
# 2. Calcule l'énergie actuelle
|
| 173 |
+
energy = energy_landscape.compute_energy(
|
| 174 |
+
current_state,
|
| 175 |
+
neighbor_vectors,
|
| 176 |
+
neighbor_ids,
|
| 177 |
+
)
|
| 178 |
+
|
| 179 |
+
result.energy_trajectory.append(energy)
|
| 180 |
+
result.neighbor_trajectory.append(neighbors_info)
|
| 181 |
+
|
| 182 |
+
# 3. Détermine les flips optimaux
|
| 183 |
+
deltas = energy_landscape.get_bit_flip_deltas(
|
| 184 |
+
current_state,
|
| 185 |
+
neighbor_vectors,
|
| 186 |
+
neighbor_ids,
|
| 187 |
+
)
|
| 188 |
+
|
| 189 |
+
# Sélectionne les flips qui réduisent l'énergie
|
| 190 |
+
# avec bruit thermique pour exploration
|
| 191 |
+
flip_probs = np.exp(-deltas / max(temperature, 0.01))
|
| 192 |
+
flip_probs = flip_probs / np.sum(flip_probs)
|
| 193 |
+
|
| 194 |
+
# Choix déterministe + stochastique
|
| 195 |
+
n_candidates = max(1, int(VECTOR_SIZE * 0.005)) # ~20 bits
|
| 196 |
+
top_candidates = np.argsort(-flip_probs)[:n_candidates * 2]
|
| 197 |
+
|
| 198 |
+
# Favorise les flips qui réduisent l'énergie
|
| 199 |
+
beneficial = deltas[top_candidates] < 0
|
| 200 |
+
if np.any(beneficial):
|
| 201 |
+
# Fait tous les flips bénéfiques avec probabilité selon température
|
| 202 |
+
selected = top_candidates[
|
| 203 |
+
np.random.random(len(top_candidates)) < flip_probs[top_candidates]
|
| 204 |
+
]
|
| 205 |
+
else:
|
| 206 |
+
# Échappement local : flip aléatoire contrôlé
|
| 207 |
+
selected = np.random.choice(
|
| 208 |
+
VECTOR_SIZE,
|
| 209 |
+
size=max(1, int(n_candidates * temperature)),
|
| 210 |
+
replace=False,
|
| 211 |
+
p=flip_probs
|
| 212 |
+
)
|
| 213 |
+
|
| 214 |
+
if len(selected) > 0:
|
| 215 |
+
new_state = current_state.copy()
|
| 216 |
+
new_state[selected] = 1 - new_state[selected]
|
| 217 |
+
|
| 218 |
+
# Calcule la nouvelle énergie
|
| 219 |
+
new_energy = energy_landscape.compute_energy(
|
| 220 |
+
new_state,
|
| 221 |
+
neighbor_vectors,
|
| 222 |
+
neighbor_ids,
|
| 223 |
+
)
|
| 224 |
+
|
| 225 |
+
# Acceptation Metropolis-Hastings
|
| 226 |
+
delta_e = new_energy - energy
|
| 227 |
+
if delta_e < 0 or np.random.random() < np.exp(-delta_e / max(temperature, 0.01)):
|
| 228 |
+
current_state = new_state
|
| 229 |
+
energy = new_energy
|
| 230 |
+
|
| 231 |
+
state_trajectory.append(current_state.copy())
|
| 232 |
+
|
| 233 |
+
# 4. Apprentissage en ligne
|
| 234 |
+
if self.online_learning:
|
| 235 |
+
learning_events = self._online_learning_step(
|
| 236 |
+
current_state,
|
| 237 |
+
neighbor_vectors,
|
| 238 |
+
neighbor_ids,
|
| 239 |
+
neighbor_indices,
|
| 240 |
+
energy,
|
| 241 |
+
iteration,
|
| 242 |
+
memory_table,
|
| 243 |
+
energy_landscape,
|
| 244 |
+
)
|
| 245 |
+
result.learning_events.extend(learning_events)
|
| 246 |
+
|
| 247 |
+
# 5. Détection périodique de motifs pour abstraction
|
| 248 |
+
if iteration > 0 and iteration % self.pattern_detection_interval == 0:
|
| 249 |
+
patterns = memory_table.detect_frequent_patterns(
|
| 250 |
+
[st for st in state_trajectory[-self.pattern_detection_interval:]],
|
| 251 |
+
min_frequency=3
|
| 252 |
+
)
|
| 253 |
+
|
| 254 |
+
for pattern in patterns:
|
| 255 |
+
# Crée une abstraction si le pattern est fréquent
|
| 256 |
+
new_id = memory_table.create_vector(
|
| 257 |
+
context=pattern,
|
| 258 |
+
abstraction_level=1,
|
| 259 |
+
)
|
| 260 |
+
result.creation_events.append({
|
| 261 |
+
'type': 'abstraction',
|
| 262 |
+
'id': new_id,
|
| 263 |
+
'iteration': iteration,
|
| 264 |
+
})
|
| 265 |
+
|
| 266 |
+
# 6. Callback externe
|
| 267 |
+
if external_callback:
|
| 268 |
+
external_callback({
|
| 269 |
+
'iteration': iteration,
|
| 270 |
+
'energy': energy,
|
| 271 |
+
'state': current_state,
|
| 272 |
+
'neighbors': neighbors_info,
|
| 273 |
+
})
|
| 274 |
+
|
| 275 |
+
# 7. Vérification convergence
|
| 276 |
+
energy_window.append(energy)
|
| 277 |
+
if len(energy_window) > self.convergence_window:
|
| 278 |
+
energy_window.pop(0)
|
| 279 |
+
|
| 280 |
+
if len(energy_window) >= self.convergence_window:
|
| 281 |
+
energy_std = np.std(energy_window)
|
| 282 |
+
energy_mean = np.mean(energy_window)
|
| 283 |
+
if energy_std / max(abs(energy_mean), 1.0) < self.early_stop_threshold:
|
| 284 |
+
result.converged = True
|
| 285 |
+
break
|
| 286 |
+
|
| 287 |
+
# Refroidissement
|
| 288 |
+
temperature *= self.cooling_rate
|
| 289 |
+
prev_energy = energy
|
| 290 |
+
|
| 291 |
+
# Inférence terminée
|
| 292 |
+
result.final_state = current_state.copy()
|
| 293 |
+
result.n_iterations = iteration + 1
|
| 294 |
+
|
| 295 |
+
# Apprentissage post-inférence : renforce les associations
|
| 296 |
+
# si l'inférence a convergé vers un état stable
|
| 297 |
+
if self.online_learning and result.converged:
|
| 298 |
+
self._post_inference_learning(
|
| 299 |
+
result,
|
| 300 |
+
memory_table,
|
| 301 |
+
energy_landscape,
|
| 302 |
+
router,
|
| 303 |
+
)
|
| 304 |
+
|
| 305 |
+
t1 = time.time()
|
| 306 |
+
result.execution_time_ms = (t1 - t0) * 1000
|
| 307 |
+
|
| 308 |
+
# Stats
|
| 309 |
+
self.total_inferences += 1
|
| 310 |
+
self.total_iterations += result.n_iterations
|
| 311 |
+
if result.converged:
|
| 312 |
+
self.total_converged += 1
|
| 313 |
+
self.avg_inference_time_ms = (
|
| 314 |
+
self.avg_inference_time_ms * (self.total_inferences - 1) + result.execution_time_ms
|
| 315 |
+
) / self.total_inferences
|
| 316 |
+
|
| 317 |
+
return result
|
| 318 |
+
|
| 319 |
+
def _online_learning_step(
|
| 320 |
+
self,
|
| 321 |
+
state: np.ndarray,
|
| 322 |
+
neighbor_vectors: np.ndarray,
|
| 323 |
+
neighbor_ids: List[int],
|
| 324 |
+
neighbor_indices: List[int],
|
| 325 |
+
energy: float,
|
| 326 |
+
iteration: int,
|
| 327 |
+
memory_table,
|
| 328 |
+
energy_landscape,
|
| 329 |
+
) -> List[Dict]:
|
| 330 |
+
"""
|
| 331 |
+
Effectue un pas d'apprentissage pendant l'inférence.
|
| 332 |
+
Mises à jour locales uniquement.
|
| 333 |
+
|
| 334 |
+
Returns:
|
| 335 |
+
Liste d'événements d'apprentissage
|
| 336 |
+
"""
|
| 337 |
+
events = []
|
| 338 |
+
|
| 339 |
+
# Met à jour les métadonnées des voisins
|
| 340 |
+
for idx in neighbor_indices:
|
| 341 |
+
if memory_table.active_mask[idx]:
|
| 342 |
+
meta = memory_table.metadata[idx]
|
| 343 |
+
meta.record_access(memory_table.time_step, energy)
|
| 344 |
+
|
| 345 |
+
# Met à jour le paysage d'énergie
|
| 346 |
+
is_stable = iteration > 5 and len(energy_landscape.energy_history) > 10
|
| 347 |
+
energy_landscape.update_from_state(
|
| 348 |
+
state,
|
| 349 |
+
neighbor_ids,
|
| 350 |
+
energy,
|
| 351 |
+
is_stable=is_stable,
|
| 352 |
+
)
|
| 353 |
+
|
| 354 |
+
# Met à jour les coactivations entre voisins
|
| 355 |
+
for i, idx1 in enumerate(neighbor_indices):
|
| 356 |
+
for idx2 in neighbor_indices[i+1:]:
|
| 357 |
+
if memory_table.active_mask[idx1] and memory_table.active_mask[idx2]:
|
| 358 |
+
id1 = memory_table.metadata[idx1].id
|
| 359 |
+
id2 = memory_table.metadata[idx2].id
|
| 360 |
+
# Renforce l'association si coactivation fréquente
|
| 361 |
+
strength = 1.0 / (1.0 + energy / 1000.0)
|
| 362 |
+
memory_table.metadata[idx1].update_coactivation(id2, strength)
|
| 363 |
+
memory_table.metadata[idx2].update_coactivation(id1, strength)
|
| 364 |
+
|
| 365 |
+
# Crée un nouveau vecteur si l'état est suffisamment différent
|
| 366 |
+
# de tous les voisins (configuration récurrente ou nouvelle)
|
| 367 |
+
if iteration > 3:
|
| 368 |
+
min_neighbor_dist = min([
|
| 369 |
+
float(np.sum(state != memory_table.vectors[idx]))
|
| 370 |
+
for idx in neighbor_indices
|
| 371 |
+
if memory_table.active_mask[idx]
|
| 372 |
+
]) if neighbor_indices else float('inf')
|
| 373 |
+
|
| 374 |
+
if min_neighbor_dist > memory_table.creation_threshold:
|
| 375 |
+
# Nouvelle configuration intéressante
|
| 376 |
+
new_id = memory_table.create_vector(context=state)
|
| 377 |
+
events.append({
|
| 378 |
+
'type': 'creation',
|
| 379 |
+
'id': new_id,
|
| 380 |
+
'reason': 'novel_pattern',
|
| 381 |
+
'distance': float(min_neighbor_dist),
|
| 382 |
+
'iteration': iteration,
|
| 383 |
+
})
|
| 384 |
+
|
| 385 |
+
return events
|
| 386 |
+
|
| 387 |
+
def _post_inference_learning(
|
| 388 |
+
self,
|
| 389 |
+
result: InferenceResult,
|
| 390 |
+
memory_table,
|
| 391 |
+
energy_landscape,
|
| 392 |
+
router,
|
| 393 |
+
):
|
| 394 |
+
"""
|
| 395 |
+
Apprentissage après convergence.
|
| 396 |
+
Renforce les associations dans la trajectoire de basse énergie.
|
| 397 |
+
"""
|
| 398 |
+
if len(result.neighbor_trajectory) < 3:
|
| 399 |
+
return
|
| 400 |
+
|
| 401 |
+
# Identifie la phase de basse énergie
|
| 402 |
+
energies = np.array(result.energy_trajectory)
|
| 403 |
+
min_energy_idx = int(np.argmin(energies))
|
| 404 |
+
|
| 405 |
+
# Les voisins à ce point sont "la réponse"
|
| 406 |
+
if min_energy_idx < len(result.neighbor_trajectory):
|
| 407 |
+
stable_neighbors = result.neighbor_trajectory[min_energy_idx]
|
| 408 |
+
stable_ids = [nid for nid, _ in stable_neighbors]
|
| 409 |
+
|
| 410 |
+
# Renforce les associations entre voisins stables
|
| 411 |
+
for i, id1 in enumerate(stable_ids):
|
| 412 |
+
for id2 in stable_ids[i+1:]:
|
| 413 |
+
pair = tuple(sorted((id1, id2)))
|
| 414 |
+
current = energy_landscape.associations.get(pair, 0.0)
|
| 415 |
+
energy_landscape.associations[pair] = min(
|
| 416 |
+
1.0,
|
| 417 |
+
current + self.learning_rate * 2.0
|
| 418 |
+
)
|
| 419 |
+
|
| 420 |
+
# Met à jour le routeur avec les nouvelles routes apprises
|
| 421 |
+
final_state = result.final_state
|
| 422 |
+
final_packed = router.pack_bits_to_uint64(final_state) if hasattr(router, 'pack_bits_to_uint64') else None
|
| 423 |
+
if final_packed is not None:
|
| 424 |
+
ph = router._pattern_hash(final_packed) if hasattr(router, '_pattern_hash') else None
|
| 425 |
+
if ph is not None:
|
| 426 |
+
router.route_cache[ph] = [
|
| 427 |
+
(nid, 1.0 / (1.0 + dist))
|
| 428 |
+
for nid, dist in stable_neighbors
|
| 429 |
+
]
|
| 430 |
+
|
| 431 |
+
def get_stats(self) -> Dict:
|
| 432 |
+
convergence_rate = (
|
| 433 |
+
self.total_converged / self.total_inferences
|
| 434 |
+
if self.total_inferences > 0 else 0.0
|
| 435 |
+
)
|
| 436 |
+
avg_iterations = (
|
| 437 |
+
self.total_iterations / self.total_inferences
|
| 438 |
+
if self.total_inferences > 0 else 0.0
|
| 439 |
+
)
|
| 440 |
+
|
| 441 |
+
return {
|
| 442 |
+
'total_inferences': self.total_inferences,
|
| 443 |
+
'convergence_rate': convergence_rate,
|
| 444 |
+
'avg_iterations': avg_iterations,
|
| 445 |
+
'avg_inference_time_ms': self.avg_inference_time_ms,
|
| 446 |
+
}
|