Transient chaotic neural network with negative self-feedback memory for continuous optimisation problems

In this paper we present a transient chaotic neural network model which incorporates negative self-feedback memory. This allows the chaotic driving term in the neural network model the ability to adapt to the energy state of the network and hence drive the system towards the global minimum avoiding the issue of being trapped in local minima. The model framework also reduces the number of parameters that require tuning to the particular optimisation problem since the self-feedback bias, I0, is no longer a tunable parameter and instead evolves with the neurodynamics. We apply the model on a continuous optimisation problem, in doing so we outline the procedure for mapping the continuous energy function onto the neuron state space and embedding the energy function into the network equations. We compare performance with the Chen and Aihara model and confirm that the self-feedback memory model does outperform the Chen & Aihara model in terms of robustness, i.e., finds the global minimum for all initial values of the self-feedback bias, I0, and still has comparable convergence rate. We also perform a parameter investigation and demonstrate that the input scaling parameter, α, can induce non-transient chaos even in the non-chaotic Hopfield network, highlighting the importance of this parameter for problem specific considerations.

Citation

Rodden, E., Gascoyne, A., Naughton, L., Brennan, J., Parkes, A. (2024). Transient Chaotic Neural Network with Negative Self-feedback Memory for Continuous Optimisation Problems. In: Arai, K. (eds) Proceedings of the Future Technologies Conference (FTC) 2024, Volume 1. FTC 2024. Lecture Notes in Networks and Systems, vol 1154. Springer, Cham. https://doi.org/10.1007/978-3-031-73110-5_19

Type

Conference contribution

Language

en

Description

This is an author's accepted manuscript of an article published by Springer Nature in Proceedings of the Future Technologies Conference (FTC) 2024, Volume 1, available online: https://doi.org/10.1007/978-3-031-73110-5_19 For re-use please see Springer's terms and conditions.

Series/Report no.

Lecture Notes in Networks and Systems, 1154

ISSN

2367-3370

EISSN

2367-3389

Collections

Faculty of Science and Engineering