With the emergence of advanced techniques in the field of artificial intelligence, risk management in the financial sector has undergone significant transformation. This paper proposes a deep learning–based approach for risk modeling using Bidirectional Long Short-Term Memory (BiLSTM) networks, adapted for tabular data by excluding explicit temporal dependencies. The model is tailored to support accurate decision-making in financial risk assessment. One notable component of this study is the use of automated hyperparameter optimization (HPO) methods, which further enhance the model’s overall effectiveness. These tuning strategies yield models that are less complex, faster to train, and capable of adapting to dynamic data environments, making them suitable for integration into automated credit approval systems. The model was evaluated against baseline approaches, demonstrating improved predictive performance across key evaluation metrics, with statistical significance confirmed by McNemar’s test.
Air pollution, particularly the concentration of particulate matter ($\mathbf{P M}_{\mathbf{1 0}}$), poses significant risks to human health and the environment. In this study, we employed the FB Prophet forecasting model to predict PM10 levels in Sarajevo and applied SHAP to enhance model interpretability. Using historical meteorological data and PM10 concentration records, we evaluated the model’s performance across three prediction horizons: 30, 60, and 90 days ahead. SHAP analysis identified the key meteorological drivers influencing $\mathbf{P M}_{10}$ concentrations. Accurate long-horizon predictions can support timely planning and decision-making, while also enhancing our understanding of air pollution dynamics in Sarajevo and providing valuable guidance for environmental management and public health strategies.
Wrist-worn photoplethysmography (PPG) offers an unobtrusive biometric modality, yet state-of-the-art models rarely address the memory and runtime constraints of embedded hardware. We conduct a systematic benchmark of CNN, LSTM, Transformer, and hybrid architectures on three public datasets, UBFC-Phys, ECSMP, and EMOGNITION, encompassing 166 unique subjects and over 26,000 PPG windows. Models are evaluated under a unified preprocessing pipeline and then converted to TensorFlow Lite to gauge deployability. A lightweight LSTM-Transformer architecture achieves $\mathbf{9 9 . 1 4 \%}$ accuracy and a global Equal Error Rate of $\mathbf{0 . 1 9 \%}$. After TFLite conversion, accuracy drops by less than $0.1 \%$ while model size shrinks from 730 kB to under 100 kB. Inference latency on a representative resource-constrained CPU ranged between 13-50 s per 10s window using single-threaded TFLite; we therefore outline quantization and hardware-aware optimization as future work.
In this paper, we analyze the secrecy outage performance of the classical Wyner’s model, where both the legitimate receiver and the eavesdropper experience gamma-shadowed two-wave with diffuse power (GS-TWDP) composite fading. We derive expressions for the lower bound of the secrecy outage probability (SOP) and the probability of strictly positive secrecy capacity (SPSC) in terms of Meijer’s G-function, which can be efficiently implemented in MATLAB® and MATHEMATICA®. The derived expressions are validated through Monte Carlo simulations and used to perform detailed analysis of the impact of shadowing and multipath severity on secrecy outage performance in channels with composite fading.
Manipulating deformable objects such as textiles remains a significant challenge in robotics due to their complex dynamics, unpredictable configurations, and high-dimensional state space. In this paper, we present an integrated planning and execution framework for robotic cloth manipulation that combines symbolic task planning, physics-based simulation, and vision-guided real-world control. High-level plans are generated using Planning Domain Definition Language (PDDL) and translated into low-level primitive actions executed on a Franka Emika Panda robot. To enable perceptual grounding, we employ a vision-based pipeline using an adapted CeDiRNet model for cloth corner detection and grasp point estimation. The system is first validated in IsaacSim using a particle-based deformable cloth model and then transferred to a real-world setup with consistent performance. Our results demonstrate successful Sim2Real transfer of high-level plans, enabling reliable execution of folding tasks in both simulated and physical environments.
