Slope failures represent a significant concern due to their potential to cause substantial loss of life
and property. These failures can occur for a variety of reasons, including deforestation, heavy rainfall, and human activities such as construction and mining. To effectively mitigate such risks, it is crucial to understand the mechanisms underlying slope failure and their contributing factors.
A study conducted by Koopialipoor et al. at the Department of Mining and Metallurgy, Amirkabir University of Technology in Tehran, Iran, focuses on predicting slope stability under both dynamic and static conditions. The research employs hybrid models that integrate artificial neural networks (ANN) with various optimization techniques, namely artificial bee colony (ABC), genetic algorithm (GA), imperialist competitive algorithm, and particle swarm optimization (PSO). The primary objective of this study is to predict the factor of safety (FOS) using these advanced hybrid methodologies.
In past studies, researchers commonly employed traditional techniques such as limit equilibrium and various numerical methods to analyze factors affecting stability and predict the factor of safety (FOS). These conventional methods are often intricate and necessitate extensive repetitive calculations, which can be time-consuming and resource-intensive. Recognizing these challenges, researchers have progressively shifted towards implementing machine learning techniques, including Artificial Neural Networks (ANN), Support Vector Machines (SVM), neuro-fuzzy systems, and Least Squares Support Vector Machines (LSSVM).
Among these advanced methodologies, Artificial Neural Networks (ANN) have emerged as the most
frequently utilized technique for FOS prediction, owing in large part to their ability to model complex relationships within data. However, ANN does come with its own set of limitations, such as a propensity to become trapped in local minima during the training process and a relatively slow learning speed which
can hinder performance in some scenarios.
With the rapid advancement of convolutional algorithms, researchers have found that exploring weights and biases within ANN networks has become more efficient and effective. In light of these developments, the author developed four innovative hybrid models—Particle Swarm Optimization-ANN (PSO-ANN), Artificial Bee Colony-ANN (ABC-ANN), Genetic Algorithm-ANN (GA-ANN), and Imperialist Competitive Algorithm-ANN (ICA-ANN)—specifically designed for predicting the factor of safety. These hybrid models aim to enhance the predictive capabilities of traditional ANN by integrating optimization techniques, thereby improving performance and efficiency in FOS assessments.
The study utilized a dataset of 699 samples, including input features such as friction angle, cohesion, slope gradient, and peak ground acceleration (PGA) for predicting the factor of safety (FOS). The PGA values were used to simulate seismic activity, ranging from 0 to 0.4, where a value of 0 represents static conditions and 0.4 represents high dynamic conditions. The simulation was conducted using Geostudio, with all models positioned on bedrock to ensure rigid behavior. The dataset was divided into an 80:20 ratio, representing the training and testing sets, respectively. The artificial neural network (ANN) architecture used for all hybrid models consisted of 5 input nodes, 8 hidden nodes, and 1 output node
(5x8x1).
To achieve optimal efficiency in model performance, it is essential to fine-tune hyper-parameter settings. The parameters were refined through a systematic trial-and-error approach. For the ICA-ANN model, the optimal hyper-parameters identified are: N(country) = 300, N(decade) = 100, and N(imp) = 5, which yielded improved performance outcomes. In the case of the PSO-ANN model, the ideal hyper-parameters consist of an inertia weight of 0.25, coefficients for the velocity equation set at 2, and a swarm size of 500. For the GA-ANN model, the optimal settings are as follows: a recombination percentage of 9%, a mutation probability of 25%, a maximum generation limit of 500, and a population size of 250. Lastly, for the ABC-ANN model, the optimal hyper-parameter identified is a number of bees set at 400.
The evaluation of the models was conducted using their optimal hyper-parameters and the primary metric for comparison was the R-squared value. The results were as follows: the ICA-ANN model achieved an R-squared value of 0.969, while the PSO-ANN model performed slightly better with an R-squared value of 0.980. In contrast, the GA-ANN model recorded a lower R-squared of 0.920, and the ABC-ANN model demonstrated an R-squared value of 0.957. These findings indicate that the PSO-ANN not only outperformed the other models in terms of predictive accuracy but also established itself as a promising hybrid model in this field of study. The superior performance of PSO-ANN suggests its potential for further applications and research advancements.
Reference
Koopialipoor, M., Jahed Armaghani, D., Hedayat, A., Marto, A., & Gordan, B. (2019). Applying various hybrid intelligent systems to evaluate and predict slope stability under static and dynamic conditions. Soft Computing, 23, 5913-5929.
Machine Learning Algorithms for Crypto Predictions – stta-consulting.com. https://stta-consulting.com/machine-learning-algorithms-for-crypto-predictions/