Adaptive control with LSTM augmentation: theory and human-in-the-loop validation

Abstract

This thesis presents a novel adaptive control architecture that provides dramatically better transient response performance compared to conventional adaptive control methods. This is accomplished by the synergistic employment of a traditional Adaptive Neural Network (ANN) controller and a Long Short-Term Memory (LSTM) network. LSTM structures can take advantage of the dependencies in an input sequence, which helps predict uncertainty. We introduce a training approach through which the LSTM network learns to compensate for the deficiencies of the ANN controller in a closed-loop setting. This improves the system’s transient response and allows the controller to react to unexpected events quickly. This study also investigates the human-in-the-loop performance of the proposed control framework. Although the LSTM-augmented control method drastically improves the transient response, especially in the presence of significant and rapid uncertainty changes, its interactions with a human operator must be analyzed to ensure safe operation. First, a human pilot model is used to investigate the overall system’s behavior and explore the controller’s performance for a reference tracking task. Then, human-in-the-loop experiments are conducted to analyze how the system responds in the presence of an actual human operator in the loop. Through careful simulation studies, we demonstrate that this architecture improves the estimation accuracy on a diverse set of uncertainties. The overall system’s stability is analyzed via a rigorous Lyapunov analysis, and the proposed method is shown to be highly effective, as demonstrated through numerical simulations and human-in-the-loop experiments.