Deakin University
Browse

File(s) under permanent embargo

Neuroevolution-based autonomous robot navigation: A comparative study

Version 2 2024-06-04, 02:23
Version 1 2020-06-12, 15:32
journal contribution
posted on 2024-06-04, 02:23 authored by SMJ Jalali, S Ahmadian, Abbas KhosraviAbbas Khosravi, S Mirjalili, MR Mahmoudi, S Nahavandi
The field of neuroevolution has achieved much attention in recent years from both academia and industry. Numerous papers have reported its successful applications in different fields ranging from medical domain to autonomous systems. However, it is not clear which evolutionary optimization techniques lead to the best results. In this paper, multilayer perceptron (MLP) neural networks (NNs) are trained and optimized using four advanced bio-inspired evolutionary algorithms (EA). The algorithms are Multi-Verse Optimizer (MVO), Moth-flame optimization (MFO), Cuckoo Search (CS) and Particle Swarm Optimization (PSO). Each algorithm is equipped with two operators: evolutionary population dynamics and mutation, which impact on exploration and exploitation. Optimized MLPs are then used for the navigation of an autonomous robot. Accuracy and area under the curve metrics are used for the evaluation and comparison metrics. Moreover, two well-regarded gradient descent algorithms including Back propagation (BP) and Levenberg Marquardt (LM) are utilized to validate the results obtained by evolutionary-based MLP trainers. It is observed that MLPs developed using MFO are the most robust ones among MLPs trained using other evolutionary and gradient descent algorithms.

History

Journal

Cognitive Systems Research

Volume

62

Pagination

35-43

Location

Amsterdam, The Netherlands

ISSN

2214-4366

eISSN

1389-0417

Language

English

Publication classification

C1 Refereed article in a scholarly journal

Publisher

ELSEVIER