Synthesizing Neural Network Controllers with Closed-Loop Dissipativity Guarantees
Researchers have developed a novel method for creating neural network controllers that guarantee system stability and performance bounds. This approach utilizes integral quadratic constraints (IQCs) to model uncertainties in both the plant and the neural network's activation functions. The method formulates a linear matrix inequality (LMI) to synthesize controllers, which is then employed in a projection-based training process to ensure these dissipativity guarantees are met. The effectiveness of this technique has been demonstrated through numerical examples involving an inverted pendulum and a flexible rod system. AI
IMPACT Introduces a formal method for ensuring stability and performance in AI-controlled systems, potentially increasing trust and adoption in safety-critical applications.