Input-to-state stable continuous time recurrent neural networks for transient circuit simulation

Yang, Alan

Permalink

https://hdl.handle.net/2142/113934 Copy

Description

Title

Input-to-state stable continuous time recurrent neural networks for transient circuit simulation

Author(s)

Yang, Alan

Issue Date

2021-12-10

Director of Research (if dissertation) or Advisor (if thesis)

• Rosenbaum, Elyse
• Raginsky, Maxim

Department of Study

Electrical & Computer Eng

Discipline

Electrical & Computer Engr

Degree Granting Institution

University of Illinois at Urbana-Champaign

Degree Name

M.S.

Degree Level

Thesis

Date of Ingest

2022-04-29T21:35:55Z

Keyword(s)

• Input to state stability
• neural ODE
• learning dynamics
• circuit simulation
• recurrent neural network

Abstract

This thesis proposes a learning approach for continuous-time recurrent neural network (CTRNN) architectures with zero or one hidden layers that guarantees input-to-state stability (ISS). We propose a model parametrization that guarantees the ISS property with respect to a Lur'e-type ISS Lyapunov function that is learned in conjunction with the model parameters. Our stability constraints impose a physical prior on the learned model, and in some cases improve the convergence of model training. The proposed CTRNN models are used to learn fast-to-simulate transient behavioral models for electronic circuits that can be implemented in the Verilog-A analog behavioral modeling language and simulated in commercial circuit simulators. The proposed CTRNNs are used to learn models of a common-source amplifier and a continuous-time linear equalizer that accurately reproduce the original circuits' behavior when interconnected in circuit configurations not encountered during model training.

Graduation Semester

2021-12

Type of Resource

Thesis

Permalink

http://hdl.handle.net/2142/113934

Copyright and License Information

Copyright 2021 Alan Yang

Owning Collections