Vadlamani Ravi

The covariance matrix adaptation evolution strategy (CMA-ES) is a population-basedblack-box optimization approach that has been used in many applications. However, it hasits own limitations, especially in terms of scalability, global exploration, poor handlingof multimodal and noisy landscapes, etc. In this work, we use opposition-based learning(OBL) to address some of the limitations of CMA-ES.We validate our proposed approachon a diverse set of benchmark test functions at different dimensions. We also establishthe statistical significance of results using the Wilcoxon signed-rank test. Additionally,we evaluate the performance of the proposed approach on the task of training neuralnetwork policies in reinforcement learning for multiple control problems. The experimentalresults demonstrate that incorporating OBL in CMA-ES leads to an improved and stableperformance on various continuous optimization and reinforcement learning tasks.

Anomaly detection is a critical component of modern financial systems, where identifying fraudulent transactions and other irregular activities is essential for maintaining security and reliability. However, such tasks are often characterized by highly imbalanced datasets containing little to no anomalous instances during training. One-Class Classification (OCC) methods, such as the Least-Squares One-Class Support Vector Machine (LS-OCSVM), are well suited to these settings, as they learn a representation of normal behaviour using only non-anomalous data. However, financial datasets often exhibit complex, non-linear relationships that can be difficult for classical kernel methods to model effectively.,To address this challenge, this work proposes the integration of quantum kernel methods and the LS-OCSVM framework for anomaly detection. By employing quantum feature maps to construct kernel functions, the proposed Quantum LS-OCSVM aims to capture complex decision boundaries that may be difficult to obtain in classical feature spaces. The performance of the quantum and the classical LS-OCSVM models is compared across multiple datasets taken from finance and cyber security.,In addition, to improve the interpretability of the resulting models, decision trees are trained on the outputs of both the LS-OCSVM models to generate human-interpretable decision rules that approximate their learned decision boundaries. This aspect provides insight into the features influencing model predictions while preserving the anomaly detection capabilities of the underlying classifiers. The proposed framework combines quantum-enhanced anomaly detection with interpretable rule-based explanations, contributing toward more effective and trustworthy financial anomaly detection systems.