Disturbance-Aware Neuro-Optimal System Control Using Generative Adversarial Control Networks.

Area of Science:

• Control Theory
• Machine Learning
• Artificial Intelligence

Background:

• Real-world systems face unavoidable disturbances impacting state and output.
• Robust control methods ensure stability but compromise optimality.
• Existing approaches struggle with unknown future disturbances.

Purpose of the Study:

• To propose Generative Adversarial Control Networks (GACNs) for robust and optimal control.
• To train controllers using optimal demonstrations without prior knowledge of disturbances.
• To explicitly determine the objective function in disturbed systems.

Main Methods:

• Formulating optimal control problems in the presence of disturbance.
• Training a controller (generator) via demonstrations of an optimal controller.
• Designing a joint loss function combining adversarial and least square losses.

Main Results:

• GACNs achieve neuro-optimal solutions for systems with unknown disturbances.
• The proposed method outperforms existing control techniques in simulations.
• Controllers trained with GACNs demonstrate enhanced stability and performance.

Conclusions:

• GACNs offer a novel approach to robust and optimal control under disturbance.
• The method effectively learns optimal control policies from demonstrations.
• GACNs provide a promising direction for intelligent control systems.