| """ |
| Paysage d'Énergie Apprenant |
| |
| La fonction d'énergie évalue la cohérence d'un état du système. |
| Elle doit être : |
| - Locale : les mises à jour ne dépendent que des voisins |
| - Apprenante : s'ajuste avec l'expérience |
| - Discriminative : basse énergie = cohérent, haute = incohérent |
| """ |
|
|
| import numpy as np |
| from numba import njit, prange |
| from typing import Dict, List, Tuple, Optional, Set |
| import logging |
|
|
| logger = logging.getLogger(__name__) |
|
|
| VECTOR_SIZE = 4096 |
|
|
|
|
| def compute_local_hamming_energy_fast( |
| state: np.ndarray, |
| neighbors: np.ndarray, |
| weights: np.ndarray, |
| ) -> float: |
| """ |
| Énergie locale de Hamming : somme pondérée des distances aux voisins. |
| Version numpy vectorisée rapide. |
| """ |
| N = neighbors.shape[0] |
| if N == 0: |
| return 0.0 |
| |
| xor_result = state.astype(np.uint8) ^ neighbors |
| |
| dists = np.sum(xor_result, axis=1) |
| return float(np.dot(weights, dists)) / N |
|
|
|
|
| def compute_bit_flip_delta_fast( |
| state: np.ndarray, |
| neighbors: np.ndarray, |
| weights: np.ndarray, |
| biases: np.ndarray, |
| ) -> np.ndarray: |
| """ |
| Calcule le delta d'énergie pour chaque flip de bit possible. |
| Version numpy vectorisée. |
| """ |
| N = neighbors.shape[0] |
| if N == 0: |
| return biases.copy() |
| |
| |
| |
| |
| |
| state_expanded = state[np.newaxis, :] |
| |
| xor_current = state_expanded ^ neighbors |
| |
| flipped = 1 - state_expanded |
| xor_flipped = flipped ^ neighbors |
| |
| |
| delta_xor = xor_flipped.astype(np.float64) - xor_current.astype(np.float64) |
| |
| weights_col = weights[:, np.newaxis] |
| weighted_delta = delta_xor * weights_col |
| |
| deltas = np.sum(weighted_delta, axis=0) |
| deltas += biases * (2.0 * (1 - state.astype(np.float64)) - 1.0) |
| |
| return deltas |
|
|
|
|
| class EnergyLandscape: |
| """ |
| Paysage d'énergie apprenant avec mises à jour locales. |
| |
| Components: |
| - hamming_weight: poids de la distance de Hamming |
| - association_weight: poids des associations Hebbian |
| - structure_weight: poids de la cohérence structurelle (binding) |
| - biases: biais par bit (apprenant) |
| - local_weights: poids par voisin (apprenant) |
| |
| Apprentissage: |
| - Renforcement des associations dans les états de basse énergie |
| - Affaiblissement dans les états instables |
| - Ajustement des biais locaux |
| """ |
| |
| def __init__( |
| self, |
| hamming_weight: float = 1.0, |
| association_weight: float = 0.5, |
| structure_weight: float = 0.3, |
| bias_learning_rate: float = 0.01, |
| association_decay: float = 0.99, |
| energy_threshold_low: float = 500.0, |
| energy_threshold_high: float = 1500.0, |
| ): |
| self.hamming_weight = hamming_weight |
| self.association_weight = association_weight |
| self.structure_weight = structure_weight |
| self.bias_lr = bias_learning_rate |
| self.assoc_decay = association_decay |
| self.low_threshold = energy_threshold_low |
| self.high_threshold = energy_threshold_high |
| |
| |
| self.biases = np.zeros(VECTOR_SIZE, dtype=np.float64) |
| |
| |
| self.associations: Dict[Tuple[int, int], float] = {} |
| |
| |
| self.structure_weights: Dict[Tuple[int, int], float] = {} |
| |
| |
| self.energy_history: List[float] = [] |
| self.min_energy_history: List[float] = [] |
| self.stable_states: List[np.ndarray] = [] |
| |
| |
| self.n_updates = 0 |
| self.total_energy = 0.0 |
| |
| def compute_energy( |
| self, |
| state: np.ndarray, |
| neighbor_vectors: np.ndarray, |
| neighbor_ids: List[int], |
| neighbor_metadata: Optional[List] = None, |
| ) -> float: |
| """ |
| Calcule l'énergie totale d'un état. |
| |
| Args: |
| state: vecteur courant (4096,) uint8 |
| neighbor_vectors: (N, 4096) voisins actifs |
| neighbor_ids: IDs des voisins |
| neighbor_metadata: métadonnées optionnelles |
| |
| Returns: |
| énergie scalaire (plus bas = plus cohérent) |
| """ |
| N = len(neighbor_ids) if neighbor_ids else 0 |
| if N == 0: |
| return float(np.sum(self.biases * (2.0 * state - 1.0))) |
| |
| |
| weights = np.array([ |
| self.associations.get( |
| tuple(sorted((-1, nid))), 0.5 |
| ) + 0.5 |
| for nid in neighbor_ids |
| ], dtype=np.float64) |
| |
| |
| hamming_energy = compute_local_hamming_energy_fast( |
| state, neighbor_vectors, weights |
| ) |
| hamming_energy *= self.hamming_weight |
| |
| |
| association_energy = 0.0 |
| if len(neighbor_ids) > 1: |
| for i in range(len(neighbor_ids)): |
| for j in range(i+1, len(neighbor_ids)): |
| pair = tuple(sorted((neighbor_ids[i], neighbor_ids[j]))) |
| strength = self.associations.get(pair, 0.0) |
| |
| dist = np.sum(neighbor_vectors[i] != neighbor_vectors[j]) |
| association_energy += strength * dist |
| |
| association_energy *= self.association_weight |
| |
| |
| structure_energy = 0.0 |
| if len(neighbor_ids) > 1: |
| for i in range(len(neighbor_ids)): |
| for j in range(i+1, len(neighbor_ids)): |
| pair = tuple(sorted((neighbor_ids[i], neighbor_ids[j]))) |
| struct_weight = self.structure_weights.get(pair, 0.0) |
| |
| corr = np.mean(neighbor_vectors[i] == neighbor_vectors[j]) |
| structure_energy += struct_weight * (1.0 - corr) |
| |
| structure_energy *= self.structure_weight |
| |
| |
| bias_energy = float(np.sum(self.biases * (2.0 * state.astype(np.float64) - 1.0))) |
| |
| total = hamming_energy + association_energy + structure_energy + bias_energy |
| |
| self.energy_history.append(total) |
| if len(self.energy_history) > 1000: |
| self.energy_history = self.energy_history[-1000:] |
| |
| return total |
| |
| def get_bit_flip_deltas( |
| self, |
| state: np.ndarray, |
| neighbor_vectors: np.ndarray, |
| neighbor_ids: List[int], |
| ) -> np.ndarray: |
| """ |
| Calcule le delta d'énergie pour chaque bit flip possible. |
| Utilisé par l'inférence pour trouver les meilleurs flips. |
| |
| Returns: |
| deltas: (4096,) négatif = flip réduit l'énergie |
| """ |
| N = len(neighbor_ids) |
| if N == 0: |
| |
| return self.biases.copy() |
| |
| weights = np.array([ |
| self.associations.get(tuple(sorted((-1, nid))), 0.5) + 0.5 |
| for nid in neighbor_ids |
| ], dtype=np.float64) |
| |
| return compute_bit_flip_delta_fast(state, neighbor_vectors, weights, self.biases) |
| |
| def update_from_state( |
| self, |
| state: np.ndarray, |
| neighbor_ids: List[int], |
| energy: float, |
| is_stable: bool = False, |
| ): |
| """ |
| Met à jour le paysage d'énergie à partir d'un état observé. |
| Appelé après chaque itération d'inférence. |
| |
| Règles: |
| - Basse énergie stable : renforce les associations |
| - Haute énergie instable : affaiblit ou réorganise |
| """ |
| self.n_updates += 1 |
| self.total_energy += energy |
| |
| if len(neighbor_ids) < 2: |
| return |
| |
| |
| if len(self.energy_history) > 20: |
| recent_mean = np.mean(self.energy_history[-20:]) |
| recent_std = np.std(self.energy_history[-20:]) + 1e-8 |
| normalized_energy = (energy - recent_mean) / recent_std |
| else: |
| normalized_energy = 0.0 |
| |
| |
| is_coherent = normalized_energy < -0.5 or (energy < self.low_threshold and is_stable) |
| is_incoherent = normalized_energy > 0.5 or energy > self.high_threshold |
| |
| |
| for i in range(len(neighbor_ids)): |
| for j in range(i+1, len(neighbor_ids)): |
| pair = tuple(sorted((neighbor_ids[i], neighbor_ids[j]))) |
| |
| if is_coherent: |
| |
| current = self.associations.get(pair, 0.0) |
| self.associations[pair] = min(1.0, current + self.bias_lr * (1.0 - current)) |
| elif is_incoherent: |
| |
| current = self.associations.get(pair, 0.0) |
| self.associations[pair] = max(-0.5, current - self.bias_lr * 2.0) |
| |
| |
| state_float = state.astype(np.float64) |
| if is_coherent: |
| |
| self.biases += self.bias_lr * (2.0 * state_float - 1.0) * 0.1 |
| elif is_incoherent: |
| |
| self.biases -= self.bias_lr * (2.0 * state_float - 1.0) * 0.2 |
| |
| |
| self.biases = np.clip(self.biases, -2.0, 2.0) |
| |
| |
| if is_stable and is_coherent: |
| self.stable_states.append(state.copy()) |
| if len(self.stable_states) > 100: |
| self.stable_states = self.stable_states[-100:] |
| |
| |
| if self.n_updates % 100 == 0: |
| self._decay_associations() |
| |
| def update_structure_weights( |
| self, |
| bound_state: np.ndarray, |
| component_ids: List[int], |
| coherence: float, |
| ): |
| """ |
| Met à jour les poids structurels pour le binding. |
| """ |
| for i in range(len(component_ids)): |
| for j in range(i+1, len(component_ids)): |
| pair = tuple(sorted((component_ids[i], component_ids[j]))) |
| current = self.structure_weights.get(pair, 0.0) |
| |
| if coherence > 0.7: |
| self.structure_weights[pair] = min(1.0, current + self.bias_lr) |
| elif coherence < 0.3: |
| self.structure_weights[pair] = max(-0.5, current - self.bias_lr * 1.5) |
| |
| def _decay_associations(self): |
| """Décroissance périodique des associations peu utilisées.""" |
| to_remove = [] |
| for pair, strength in self.associations.items(): |
| decayed = strength * self.assoc_decay |
| if abs(decayed) < 0.01: |
| to_remove.append(pair) |
| else: |
| self.associations[pair] = decayed |
| |
| for pair in to_remove: |
| del self.associations[pair] |
| |
| def get_association_strength(self, id1: int, id2: int) -> float: |
| """Retourne la force d'association entre deux vecteurs.""" |
| pair = tuple(sorted((id1, id2))) |
| return self.associations.get(pair, 0.0) |
| |
| def get_stats(self) -> Dict: |
| n_assoc = len(self.associations) |
| n_struct = len(self.structure_weights) |
| recent_energy = np.mean(self.energy_history[-100:]) if self.energy_history else 0.0 |
| |
| return { |
| 'n_associations': n_assoc, |
| 'n_structure_weights': n_struct, |
| 'mean_bias': float(np.mean(np.abs(self.biases))), |
| 'recent_mean_energy': float(recent_energy), |
| 'n_updates': self.n_updates, |
| 'n_stable_states': len(self.stable_states), |
| } |