Integration of Sense and Control for Uncertain Systems Based on Delayed Feedback Active Inference

Area of Science:

• Robotics and Control Systems
• Computational Neuroscience
• Artificial Intelligence

Background:

• Sensor data transmission introduces time lags, causing delayed information that deviates from the current system state.
• Active inference (AIF) often overlooks input delays, leading to inaccurate state estimation and prediction errors, particularly with external disturbances.

Purpose of the Study:

• To propose a theoretical framework, delayed feedback active inference (DAIF), to improve AIF's applicability in real-world systems with time-delayed data.
• To enhance state estimation and prediction accuracy in dynamic systems by accounting for input delays.

Main Methods:

• Defined a probability model for DAIF by integrating a control distribution into the AIF model.
• Formulated DAIF free energy as the sum of quadratic state, sense, and control prediction errors.
• Introduced a predicted state derived from past states and a proportional-integral (PI)-like control mechanism based on this prediction.

Main Results:

• DAIF incorporates a predicted state and PI-like control, with gain inversely proportional to measurement accuracy variance.
• Employed a second-order inverse variance accuracy to adaptively compensate for external disturbances, replacing fixed sensory accuracy.
• Simulations on a quadrotor unmanned aerial vehicle (UAV) demonstrated DAIF's superior performance in trajectory tracking control compared to AIF.

Conclusions:

• DAIF effectively compensates for time-lagged sensor data, outperforming standard AIF in state estimation and disturbance resistance.
• The proposed DAIF framework enhances the robustness and applicability of active inference for real-time control systems.