Stabilizing the Convergence of Pixel-Based Deep Active Inference Controllers Using Adaptive Smoothing Filters.

Area of Science:

• Robotics
• Artificial Intelligence
• Computational Neuroscience

Background:

• Active inference is increasingly used for robot control due to its environmental adaptability.
• Conventional active inference controllers using gradient descent risk convergence to suboptimal solutions (local minima).

Purpose of the Study:

• To propose a novel approach using a smoothing filter to mitigate local minima in pixel-based active inference controllers.
• To enhance the convergence performance and robustness of active inference in robot control.

Main Methods:

• A smoothing filter was applied to observed, predicted, and target values within the active inference framework.
• The smoothing intensity was dynamically adjusted based on prediction and target errors to balance smoothing and gradient information.
• The method was evaluated in simulation using object tracking and robotic arm control tasks.

Main Results:

• The proposed smoothing filter approach demonstrated improved convergence performance compared to conventional active inference controllers.
• The dynamic adjustment of smoothing intensity prevented the loss of essential gradient information.

Conclusions:

• Smoothing pixel-based active inference controllers effectively reduces the risk of local minima.
• The proposed method offers a more reliable and efficient approach for robot control applications using active inference.