A study conducted by Pijush Samui et al. from VIT University, Vellore, and published in the Geotechnical and Geological Engineering Journal in April 2014, focuses on predicting Standard Penetration Test (SPT) seismic liquefaction using the Minimax Probability Machine (MPM) with a radial basis function (RBF). Liquefaction refers to the phenomenon where soil behaves like a liquid, resulting in significant structural damage. Given the complexity of the liquefaction mechanism, accurately determining its seismic implications poses a considerable challenge.
Researchers have proposed a variety of machine learning models, among which Artificial Neural Networks (ANN) and Support Vector Machines (SVM) are prominent choices. However, each of these models comes with its own set of limitations that can affect their performance and applicability in real-world scenarios. Starting with Artificial Neural Networks, one of the major criticisms is that they often operate as “black boxes.” This means that while they can process and analyze complex patterns in data, the reasoning behind their decision-making is not easily interpretable, making it challenging to understand how specific outcomes are achieved. Additionally, ANNs may struggle with generalization, which refers to the model’s ability to perform well on unseen data. They are also prone to arriving at local minima during the training process, which can limit their ability to find the best solution. Furthermore, overfitting is a common issue, where the model learns the training data too thoroughly and fails to generalize to new, unseen data effectively.
On the other hand, Support Vector Machines are known for their effectiveness in high-dimensional spaces but come with their own set of challenges. The computational complexity of SVMs is notably high due to the need for quadratic programming, which can make them less efficient, especially with larger datasets. Moreover, determining the optimal tuning parameters, such as the kernel type and regularization factor, can be a difficult and time-consuming task, requiring extensive cross-validation and experimentation to achieve optimal performance. Overall, while both ANN and SVM have their strengths and areas of application, understanding their limitations is crucial for researchers and practitioners when selecting the most suitable model for their specific needs.
In light of the limitations previously discussed, the authors have employed the Multiclass Probability Machine (MPM) in their research. Developed by Lanckriet et al., MPM operates within a probabilistic framework. This approach serves as an alternative to Support Vector Machines (SVM), emphasizing the minimization of misclassification probability to ensure robust and reliable classification boundaries, even under uncertain conditions. The primary objective is to identify a separating hyperplane that maximizes the minimum probability of correctly classifying data points, based on the assumption that the data follows a Gaussian distribution.
The study utilizes a dataset that includes parameters such as cyclic stress ratio (CSR), peak ground acceleration (PGA), standard penetration test value (N), and the status of soil. Two predictive models have
been developed: MODEL I, which incorporates N and CSR as input variables, and MODEL II, which utilizes PGA and N as inputs. The dataset comprises a total of 288 samples, comprised of 164 liquefied samples and 124 non-liquefied samples. The dataset has been partitioned into 70% for training and 30% for testing purposes. The hyperparameter (σ) was established through a methodical trial and error approach.
Based on the analysis of the accuracy metrics, Model I exhibited a σ value of 0.8, achieving an accuracy rate of 97.67%. In comparison, Model II, with a σ value of 1, demonstrated an accuracy of 96.51%.
This study applies a Modified Prediction Model (MPM) designed to evaluate the seismic liquefaction potential of soil by utilizing the N value and Peak Ground Acceleration (PGA). The model was developed using a dataset comprising 288 samples, resulting in two distinct models (MODEL I and MODEL II), both of which exhibited exceptional performance and generalization capabilities.
Notably, the MPM relies solely on PGA and the N value for its predictions, thereby eliminating the requirement for Cyclic Stress Ratio (CSR). The findings suggest that MPM-based methodologies offer a promising approach for assessing the potential for soil liquefaction.
Reference
Samui, P., & Hariharan, R. (2014). Modeling of SPT seismic liquefaction data using minimax probability machine. Geotechnical and Geological Engineering, 32, 699-703.
Data, S. L. (2015). Minimax Probability Machine.