580 California St., Suite 400
San Francisco, CA, 94104
Key research themes
1. How can neural networks enable robust and adaptive control of nonlinear dynamic systems with unknown or uncertain dynamics?
This research theme investigates neural network control methods designed to manage nonlinear dynamical systems characterized by unknown, uncertain, or changing dynamics. Such systems resist conventional control methods relying on precise analytical models. Neural networks, with their universal approximation capabilities and learning properties, are leveraged to estimate inverse dynamics, compensate unmodeled effects, and adapt controller parameters online to guarantee stability and tracking performance despite system uncertainties or disturbances. The works collectively focus on ensuring global or robust stability typically via Lyapunov-based approaches and online learning algorithms, and demonstrate effectiveness on nonlinear robot manipulators and other dynamic systems.
Key finding: Develops a neural control scheme for 3D robotic manipulators with completely unknown dynamics, using an online learning algorithm grounded in Lyapunov stability theory to guarantee global stability despite perturbations such... Read more
Key finding: Employs a three-layer recurrent neural network to estimate forward robot dynamics, controlled via online backpropagation to minimize trajectory tracking errors in nonlinear time-varying robotic manipulators. The approach... Read more
Key finding: Proposes an adaptive neural network-based output feedback controller for highly nonlinear, unstable, and underactuated systems exemplified by a two-wheel mobile robot (inverted pendulum model). The controller adapts online to... Read more
Key finding: Compares recurrent NARX and NARMA-L2 networks with feedforward neural networks for adaptive control of nonlinear discrete-time systems with unknown dynamics. The study highlights the advantage of using neural network... Read more
Key finding: Introduces the application of the SPSA algorithm for efficient online training of neural network controllers managing nonlinear systems, specifically the adaptive control of a Translational Oscillator with Rotational Actuator... Read more
2. What neural network architectures and training methods enhance control performance for nonlinear systems with large time delays and nonlinearities?
This theme examines how different neural network structures and associated training algorithms are applied to nonlinear control problems involving significant time delays, nonlinearities, and uncertainties. It focuses on architectures like recurrent networks (NARX, NARMA-L2), feedforward multilayer perceptrons, and combinations (e.g., neuro-fuzzy), alongside specialized online learning or optimization schemes such as backpropagation and SPSA that enable robust, fast adaptive control. The research covers domain-specific applications like irrigation canal flow regulation, electrical power converters, and chemical process control, emphasizing improved disturbance rejection, setpoint tracking, and system stability within these challenging environments.
Key finding: Develops a combined control system coupling a Smith Predictor (SP) with a NARX artificial neural network that models nonlinear pool dynamics to manage irrigation canal water distribution exhibiting large time delays and... Read more
Key finding: Proposes an online learning neural network controller using backpropagation for a Buck-Boost DC-DC converter to stabilize and regulate output voltage amidst transient conditions. The controller adapts parameters in real time,... Read more
Key finding: Introduces a multilayer feedforward neural network framework for PID controller gain tuning, employing backpropagation to adaptively optimize gains and stabilize unstable nonlinear systems, exemplified by a magnetic... Read more
Key finding: Provides an extensive survey of neural network applications in chemical process control systems, categorizing approaches into predictive, inverse-model-based, and adaptive control. The review highlights that multilayer... Read more
Key finding: Presents a four-layer feedforward neural network controller for autonomous mobile robot navigation in dynamic and partially unknown environments. The network inputs comprise sensor data related to obstacles and target... Read more
3. How do hybrid and neuro-fuzzy approaches integrate with neural networks to improve control of nonlinear dynamic systems?
This theme explores the integration of neural networks with fuzzy logic and other soft computing techniques to design control schemes for complex nonlinear systems. It assesses how fuzzy-recurrent high order neural networks and neuro-fuzzy inference systems extend approximation capabilities and enhance robustness, stability, and adaptive tracking performance. These hybrid methods aim to leverage interpretability from fuzzy systems and learning flexibility from neural networks to tackle control problems where analytical models are inadequate or unavailable, with guarantees on convergence and boundedness.
Key finding: Develops adaptive indirect and direct regulation control schemes combining fuzzy dynamical systems with high order neural networks (F-RHONNs) to approximate unknown nonlinear plants. The weight update laws ensure exponential... Read more
Key finding: Proposes a structurally simple neural network-based controller employing fast, explicit single-step teaching and data retrieval using neuron activation functions implementing abstract rotations for mapping kinematic data to... Read more
Control of direct current motor by using artificial neural networks in Internal model control scheme
Key finding: Applies two feedforward neural networks trained on historical data as an internal model and its inverse to implement an Internal Model Control (IMC) scheme for DC motor speed regulation. The neuro controller achieves... Read more
Key finding: Compares several artificial intelligence controllers including neural networks (PI-ANN), fuzzy logic (PI-FLC), and neuro-fuzzy (PI-NFLC) for pitch angle regulation in large wind turbines integrated with Maximum Power Point... Read more
Key finding: Implements a neural network controller for spacecraft power system regulation in low earth orbit, trained via backpropagation with the Levenberg-Marquardt algorithm. The ANN achieves near-perfect regression accuracy and mean... Read more