ePrints@IIScePrints@IISc Home | About | Browse | Latest Additions | Advanced Search | Contact | Help

On the Use of Dropouts in Neural Networks for System Identification and Control

Yadav, SM and George, K (2019) On the Use of Dropouts in Neural Networks for System Identification and Control. In: UNSPECIFIED, pp. 1374-1381.

[img] PDF
Iee_Sym_Ser_Com_Int_1374_2019.pdf - Published Version
Restricted to Registered users only

Download (703kB) | Request a copy
Official URL: https://dx.doi.org/10.1109/SSCI.2018.8628706


As universal function approximators, neural networks have been successfully used for nonlinear dynamical system identification and control. Recent developments in neural network regularisation methods include dropouts. Here, the outputs of hidden units are dropped randomly with a probability which is tuned as a hyperparameter. In this paper, we first present a scheme to conduct a rigorous analysis on the effect of changing hyperparameters in neural network learning by comparing their ability to generalise and achieve convergence. Second, we conduct an analysis with dropouts using the proposed scheme through a grid search over the dropout probabilities in each hidden layer. In the context of system identification and control, our results show that dropout is at best as good as standard back propagation in terms of the amount of data required for generalisation without adapting. This is achieved by the proposed scheme which efficiently summarises hundreds of simulations with typical examples. © 2018 IEEE.

Item Type: Conference Poster
Publication: Proceedings of the 2018 IEEE Symposium Series on Computational Intelligence, SSCI 2018
Publisher: Institute of Electrical and Electronics Engineers Inc.
Additional Information: Copyright for this article belongs to Institute of Electrical and Electronics Engineers Inc.
Keywords: Artificial intelligence; Backpropagation; Dynamical systems; Identification (control systems); Neural networks; Nonlinear dynamical systems; Nonlinear systems; Religious buildings, Adaptive Control; dropouts; Generalisation; Hyper-parameter; Hyperparameters; Neural network learning; Rigorous analysis; Universal functions, Adaptive control systems
Department/Centre: Division of Interdisciplinary Sciences > Robert Bosch Centre for Cyber Physical Systems
Date Deposited: 15 Apr 2019 05:20
Last Modified: 15 Apr 2019 05:20
URI: http://eprints.iisc.ac.in/id/eprint/62095

Actions (login required)

View Item View Item