Journal: Entropy
Article Title: Intermittent Active Inference
doi: 10.3390/e28030269
Figure Lengend Snippet: Flow Chart for the Belief Divergence Trigger Mechanism. After the planning phase, the agent has chosen a plan π * to follow (yellow plan). After applying an action a , the agent forwards its belief by applying an unscented Kalman filter (UKF). The agent then receives an observation o and updates its belief about the system state Q s using variational inference (VI). In every time step, this updated belief is compared with the agent’s belief for this time step during planning Q ^ s (orange area). Only when the Jensen-Shannon divergence (JS Div.) surpasses the provided threshold ϵ Div , the current plan is abandoned and a new planning phase is triggered (step 5).
Article Snippet: To efficiently update the agent’s belief when performing an action, we apply an Unscented Kalman Filter (UKF) which propagates normal distributions through non-linear dynamics [ , ].
Techniques: