Lyapunov-Based Real-Time and Iterative Adjustment of Deep Neural Networks

Sun, Runhan; Greene, Max L.; Le, Duc Minh; Bell, Zachary I.; Chowdhary, Girish; Dixon, Warren E.

doi:10.1109/lcsys.2021.3055454

Cited by 36 publications

(25 citation statements)

References 20 publications

Supporting

Mentioning

Contrasting

Unclassified

Order By: Relevance

“…Efforts to combine AC and ML approaches have been highlighted in several recent works, which include [47,[55][56][57][58]. Other than our earlier work in [49,50], new algorithms that combine AC and RL in continuous-time systems can be found in [59][60][61].…”

Section: Introductionmentioning

confidence: 99%

“…References [60,61] study the linear-quadratic-regulator problem and its adaptive control variants from an optimization and machine learning perspective. In [57] a reinforcement learning approach is used to determine an adaptive controller for an unknown system, while in [58] principles from adaptive control and Lyapunov analysis are used to adjust and train a deep neural network. For the second class of algorithms in discrete-time, other than our earlier work in [50][51][52], to our knowledge, similar combinations of AC and ML for accelerated performance and convergence have not been reported to-date.…”

Section: Introductionmentioning

confidence: 99%

See 1 more Smart Citation

Integration of adaptive control and reinforcement learning for real-time control and learning

Annaswamy¹,

Guha²,

Cui³

et al. 2021

Preprint

View full text Add to dashboard Cite

This paper considers the problem of real-time control and learning in dynamic systems subjected to uncertainties. Adaptive approaches are proposed to address the problem, which are combined with methods and tools in Reinforcement Learning (RL) and Machine Learning (ML). Algorithms are proposed in continuous-time that combine adaptive approaches with RL leading to online control policies that guarantee stable behavior in the presence of parametric uncertainties that occur in real-time. Algorithms are proposed in discrete-time that combine adaptive approaches proposed for parameter and output estimation and ML approaches proposed for accelerated performance that guarantee stable estimation even in the presence of time-varying regressors, and for accelerated learning of the parameters with persistent excitation. Numerical validations of all algorithms are carried out using a quadrotor landing task on a moving platform and benchmark problems in ML. All results clearly point out the advantage of the proposed integrative approaches for real-time control and learning.

show abstract

Section: Introductionmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

Integration of adaptive control and reinforcement learning for real-time control and learning

Annaswamy¹,

Guha²,

Cui³

et al. 2021

Preprint

View full text Add to dashboard Cite

show abstract

“…El único punto que queda por clarificar es como estimar los parámetros,θ, de la regresión (9). En la estructura de (9) es común implementar un algoritmo de mínimos cuadrados recursivos [6][7].…”

Section: Estimador De Parámetros Dispersounclassified

“…Finalmente, a diferencia de muchas técnicas de aprendizaje [9], el algoritmo propuesto incorpora una métrica que le permite cuantificar la precisión de los parámetros estimados. Mediante esta métrica se podrá discernir entre una fase de aprendizaje, en la que el controlador está aprendiendo el sistema, y una fase de ejecución, en la que el controlador sigue la referencia deseada.…”

Section: Introductionunclassified

Control no lineal adaptativo con identificación dispersa

Domingo-Enrich¹,

Cecilia²,

Costa‐Castelló³

2021

Xlii Jornadas De Automática : Libro De Actas

View full text Add to dashboard Cite

En este trabajo se propone un controlador con estimación de parámetros enfocada a sistemas no lineales, desconocidos y cambiantes en el tiempo. Se presenta un algoritmo de aprendizaje en línea para estimar las ecuaciones del sistema y un controlador que usa esta información para estabilizar la planta. Para ello, se han combinado un observador no lineal y un estimador de parámetros con técnicas de optimización dispersa. En el algoritmo hay dos fases diferenciadas: la fase de aprendizaje, donde se aprenden en línea las funciones desconocidas, y la fase de ejecución, donde se controla la planta con unos objetivos preestablecidos. El algoritmo propuesto es capaz de decidir de forma autónoma en qué fase se encuentra, mediante una métrica de aprendizaje. Este algoritmo se ha validado por simulación numérica en una planta roto-magnet.

show abstract

“…Inspired by [6], [7], and [10], we develop an adaptive eventtriggered distributed state observer that utilizes DNNs as a means to improve state reconstruction for an uncertain nonlinear system. Using a nonsmooth Lyapunov stability analysis, we prove that our observer is capable of UUB state reconstruction while being robust to a bounded exogenous disturbance.…”

Section: Introductionmentioning

confidence: 99%

Distributed State Estimation with Deep Neural Networks for Uncertain Nonlinear Systems under Event-Triggered Communication

Zegers¹,

Sun²,

Chowdhary³

et al. 2022

Preprint

Self Cite

View full text Add to dashboard Cite

Distributed state estimation is examined for a sensor network tasked with reconstructing a system's state through the use of a distributed and event-triggered observer. Each agent in the sensor network employs a deep neural network (DNN) to approximate the uncertain nonlinear dynamics of the system, which is trained using a multiple timescale approach. Specifically, the outer weights of each DNN are updated online using a Lyapunov-based gradient descent update law, while the inner weights and biases are trained offline using a supervised learning method and collected input-output data. The observer utilizes event-triggered communication to promote the efficient use of network resources. A nonsmooth Lyapunov analysis shows the distributed event-triggered observer has a uniformly ultimately bounded state reconstruction error. A simulation study is provided to validate the result and demonstrate the performance improvements afforded by the DNNs.

show abstract

Lyapunov-Based Real-Time and Iterative Adjustment of Deep Neural Networks

Cited by 36 publications

References 20 publications

Integration of adaptive control and reinforcement learning for real-time control and learning

Integration of adaptive control and reinforcement learning for real-time control and learning

Control no lineal adaptativo con identificación dispersa

Distributed State Estimation with Deep Neural Networks for Uncertain Nonlinear Systems under Event-Triggered Communication

Contact Info

Product

Resources

About