An augmented mountain Gazelle optimizer for efficient training of transfer learning-based artificial neural networks
Year: 2026
Type: article
Abstract: Artificial neural networks (ANNs) are prevalent for classification and regression problems from diverse fields. Gradient-based optimization algorithms, which optimize weights and biases of the ANNs during training, have high training time and low accuracy due to the limitations that come from the gradient. Metaheuristic algorithms are alternative training algorithms for ANNs as they are efficient to explore search spaces such as multi-modal and higher-dimensional. Mountain Gazelle optimizer (MGO) is a recent metaheuristic algorithm that has notable effectiveness compared to the state-of-the-art algorithms. However, it lacks in searching potential regions, sustaining diversity in the population, and preventing local optima. To address these problems, this paper proposes an augmented mountain Gazelle optimizer (AMGO) by incorporating a strategy from a recent metaheuristic, which examines more potential regions, an empirical-based update equation, and opposition-based learning (OBL). The efficiency of the AMGO algorithm was validated by the experiments on the CEC2017 test suite benchmark functions and on training ANNs for classification and regression problems on various datasets, including potato leaf disease detection. Moreover, its performance was compared with well-known and recent metaheuristic algorithms. Further, its significance was verified with the Friedman and Wilcoxon rank-sum statistical tests. The results show that the AMGO algorithm has better performance than the MGO for 23 benchmark functions and diverse datasets. Moreover, the AMGO algorithm increases the average train and test accuracies by 2.9% and 4.4%, 1.9% and 1.8%, 3.6% and 3.8%, 10% and 9.8% compared to the MGO algorithm for the classification problems. Furthermore, it decreases the train and test RMSE values by 0.0032 and 0.0031, respectively, compared to the MGO algorithm for the regression problem. These outputs illustrate the significant improvement of the AMGO algorithm over the MGO. Experimental evaluation on the CEC2017 benchmark suite and ANN training problems demonstrates that AMGO provides competitive and generally improved optimization performance compared with the original MGO and several competitive metaheuristics. The proposed framework is particularly effective for ANN weight optimization while maintaining the computational complexity of the original MGO.
Source: Scientific Reports
Authors: Pedda Nagyalla Maddaiah(Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Project administration, Resources, Software, Validation, Visualization, Writing – original draft, Writing – review & editing), Pournami Pulinthanathu Narayanan(Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Resources, Software, Validation, Visualization, Writing – original draft, Writing – review & editing), Kumar Dilip(Conceptualization, Data curation, Investigation, Methodology, Resources, Software, Validation, Visualization, Writing – original draft, Writing – review & editing), Dilip Kumar Jang Bahadur Saini(Conceptualization, Data curation, Investigation, Resources, Software, Validation, Visualization, Writing – original draft, Writing – review & editing), Biswa Mohan Sahoo(Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Project administration, Resources, Software, Validation, Visualization, Writing – original draft, Writing – review & editing), Nripendra Narayan Das(Conceptualization, Data curation, Formal analysis, Funding acquisition, Investigation, Project administration, Resources, Software, Validation, Writing – original draft, Writing – review & editing)
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $2,690
OpenAlex ID: W7233702707
