Comparative Analysis of Ant Colony and Whale Optimization Algorithms for Hyperparameter Optimization in RNN-Based Constitutive Modeling of Unstabilized Rammed Earth

The goal of this research is to develop an accurate predictive model for the stress-strain behavior of unstabilized rammed earth in compression using Recurrent Neural Networks (RNNs), where the hyperparameters are determined using Ant Colony Optimization (ACO) and Whale Optimization Algorithm (WOA). Written in Python, the NNs were trained on synthetic data consisting of 2,000 random stress-strain curves generated from probabilistic distributions of peak strain and strength. Experiments show the effectiveness of metaheuristic search: although a manually tuned baseline mean squared error of 0.00049, a two-hand (Ant Colony Optimization and Whale Optimization Algorithm) optimized model drastically outperformed prediction performance with convergence to 0.000064 and 0.000039 in terms of Mean Squared Error (MSE), respectively. The reliability of the model was demonstrated by experimental validation with various types of soil mixtures.