File size: 1,924 Bytes
7a27384 | 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 | import tensorflow as tf
class AIFStabilityLoss(tf.keras.losses.Loss):
def __init__(self, target_equilibrium_variance=1.0, penalty_weight=0.5, name="aif_stability_loss"):
super(AIFStabilityLoss, self).__init__(name=name)
self.target_variance = target_equilibrium_variance
self.penalty_weight = penalty_weight
def call(self, initial_state, next_state):
"""
Calculates loss based on how perfectly the loop stabilizes the agent.
initial_state: The agent's intelligence vector BEFORE the loop (batch_size, 16)
next_state: The agent's intelligence vector AFTER the loop (batch_size, 16)
"""
# 1. State Drift Loss: Measures the literal distance shifted during the loop
state_drift = tf.reduce_mean(tf.square(next_state - initial_state), axis=-1)
# 2. Chaos Penalty: Prevents the parameters from exploding or flatlining.
# It forces the variance of the new state to match your target equilibrium.
current_variance = tf.math.reduce_variance(next_state, axis=-1)
chaos_penalty = tf.square(current_variance - self.target_variance)
# 3. Total Closed-Loop Loss
total_loss = state_drift + (self.penalty_weight * chaos_penalty)
return tf.reduce_mean(total_loss)
loss_fn = AIFStabilityLoss(target_equilibrium_variance=1.0, penalty_weight=0.1)
# Simulating a stable transition vs an unstable, chaotic explosion
initial_intelligence = tf.random.normal((1, 16), mean=0.0, stddev=1.0)
stable_next_state = initial_intelligence + tf.random.normal((1, 16), mean=0.0, stddev=0.1)
exploded_next_state = initial_intelligence * 5.0 # Massive parameter explosion
print("Loss for Stable Loop Transition:", loss_fn(initial_intelligence, stable_next_state).numpy())
print("Loss for Chaotic Loop Explosion :", loss_fn(initial_intelligence, exploded_next_state).numpy())
|