UC Merced

In the fast-advancing domains of artificial intelligence and machine learning,the need for models capable of efficiently handling large, complex datasets is inevitable. Traditional methods such as decision trees and nearest-neighbor algorithms often struggle with computational complexity and scalability when applied to highdimensional data, especially in domains like agriculture, where large amounts of realtime data are generated from sensor networks. These challenges require the development of more efficient learning algorithms that can reduce computational overhead while maintaining predictive accuracy. To address these challenges, in this research, we propose a novel Stochastic Decision Tree (SDT) model that introduces randomness into the tree induction process, significantly reducing computational complexity while maintaining or improving accuracy. The proposed method leverages stochasticity to prioritize important features and efficiently process large datasets. The proposed Neural-SRNN model addresses class imbalance issues by using a specialized loss function (Focal Loss) to focus on hard-to-classify instances. We demonstrate the effectiveness of this approach through empirical evaluations of standard datasets. In addition, we extend the application of stochastic methods by developing a Neural-Synthetic Reduced Nearest Neighbor (Neural-SRNN) algorithm, which integrates neural networks into the SRNN framework to improve performance and interpretability in high-dimensional classification tasks. This method combines the flexibility of neural networks with the computational advantages of the SRNN model, achieving superior performance on image classification benchmarks such as MNIST and Fashion-MNIST with an expectation-maximization approach. To preserve the strength of the proposed Neural-SRNN model and improve efficiency and scalability, we proposed a two-layer Neural-SRNN model that builds on the Synthetic Reduced Nearest-Neighbor (SRNN) architecture. This model uses a modular approach, employing Mini-Convolutional Neural Networks (Mini-CNNs) in the first layer to perform class-specific feature extraction, followed by a shallow neural network for final classification. This two-layer architecture allows for efficient parallel processing, significantly reducing computational overhead while maintaining high accuracy. Moreover, it improves generalization and prevents overfitting on complex datasets such as SignMNIST and FashionMNIST. The two-layer Neural-SRNN model demonstrates superior accuracy and computational efficiency, demonstrating its scalability and adaptability to complex, high-dimensional data. The dissertation further explores the application of the Stochastic Decision tree technique to real-world agricultural problems, particularly in optimizing water usage in almond and pistachio orchards. By predicting stem water potential using aerial and ground sensor data, the SDT model improves irrigation efficiency and contributes to the development of sustainable agricultural practices. This work advances machine learning by presenting novel and efficient stochastic algorithms and demonstrating their practical applicability in critical areas such as agriculture.