Abstract

Abstract
In the field of gaming artificial intelligence, selecting the appropriate machine learning approach is essential for improving decision-making and automation. This paper examines the effectiveness of deep reinforcement learning (DRL) within interactive gaming environments, focusing on complex decision-making tasks. Utilizing the Unity engine, we conducted experiments to evaluate DRL methodologies in simulating realistic and adaptive agent behavior. A vehicle driving game is implemented, in which the goal is to reach a certain target within a small number of steps, while respecting the boundaries of the roads. Our study compares Proximal Policy Optimization (PPO) and Soft Actor-Critic (SAC) in terms of learning efficiency, decision-making accuracy, and adaptability. The results demonstrate that PPO successfully learns to reach the target, achieving higher and more stable cumulative rewards. Conversely, SAC struggles to reach the target, displaying significant variability and lower performance. These findings highlight the effectiveness of PPO in this context and indicate the need for further development, adaptation, and tuning of SAC. This research contributes to developing innovative approaches in how ML can improve how player agents adapt and react to their environments, thereby enhancing realism and dynamics in gaming experiences. Additionally, this work emphasizes the utility of using games to evolve such models, preparing them for real-world applications, namely in the field of vehicles' autonomous driving and optimal route calculation.

Abstract
This study investigates the effects of occlusions on the fairness of face recognition systems, particularly focusing on demographic biases. Using the Racial Faces in the Wild (RFW) dataset and synthetically added realistic occlusions, we evaluate their effect on the performance of face recognition models trained on the BUPT-Balanced and BUPT-GlobalFace datasets. We note increases in the dispersion of FMR, FNMR, and accuracy alongside decreases in fairness according to Equalized Odds, Demographic Parity, STD of Accuracy, and Fairness Discrepancy Rate. Additionally, we utilize a pixel attribution method to understand the importance of occlusions in model predictions, proposing a new metric, Face Occlusion Impact Ratio (FOIR), that quantifies the extent to which occlusions affect model performance across different demographic groups. Our results indicate that occlusions exacerbate existing demographic biases, with models placing higher importance on occlusions in an unequal fashion across demographics.

Abstract
The advent of large Machine Learning models and the steep increase in the demand for AI solutions occurs at the same point in time in which policies are being enacted to implement more sustainable processes in virtually every sector. This means there is a need for more, better and larger models, which require significant computational resources, while at the same time a call for a decrease in the energy spent in the processes associated to MLOps. In this paper we propose a reduced set of meta-features that can be used to characterize sets of data and their relationship with model performance. We start from a large set of 66 features, and reduce it to only 10 while maintaining the strength of this relationship. This ensures a process of meta-feature extraction and prediction of model performance that is in line with the desiderata of Frugal AI, allowing to develop more efficient ML processes.

Abstract
This paper addresses the persistent rise in motorcycle-related fatalities, even as overall road deaths decline, by introducing an adaptive Fuzzy System based on the Takagi-Sugeno model. The system evaluates parameters such as acceleration and lean angle to classify rider behavior into categories such as normal, aggressive, or dangerous, providing timely feedback aimed at promoting safer driving practices. A key component of this approach is the Local Outlier Factor (LOF) algorithm, which identifies hazardous behaviors by quantifying deviations from standard riding patterns, thereby allowing the establishment of adaptive safety thresholds. By integrating fuzzy logic, the system offers refined decision-making capabilities in complex riding conditions, enhancing active safety systems such as traction and braking controls. This work emphasizes the critical role of behavior-based insights in mitigating accidents, particularly since rider actions are a major contributing factor to motorcycle incidents.

Abstract
Accurate traffic state prediction is fundamental to Intelligent Transportation Systems, playing a critical role in optimising traffic management, improving mobility, and enhancing the efficiency of transportation networks. Traditional methods often rely on feature engineering, statistical time-series approaches, and non-parametric techniques to model the inherent complexities of traffic states, incorporating external factors such as weather conditions and accidents to refine predictions. However, the effectiveness of long-term traffic state prediction hinges on capturing spatial-temporal dependencies over extended periods. Current models face challenges in dealing with (i) high-dimensional traffic features, (ii) error accumulation for multi-step prediction, and (iii) robustness to external factors effectively. To address these challenges, this study proposes a novel model with a Dynamic Feature Embedding layer designed to transform complex data sequences into meaningful representations and a Deep Linear Projection network that refines these representations through non-linear transformations and gating mechanisms. These two features make the model more scalable when dealing with high-dimensional traffic features. The model also includes a Spatial-Temporal Positional Encoding layer to capture spatial-temporal relationships, masked multi-head attention-based encoder blocks, and a Residual Temporal Convolutional Network to process features and extract short-and long-term temporal patterns. Finally, a Time-Distributed Fully Connected Layer produces accurate traffic state predictions up to 24 timesteps into the future. The proposed architecture uses a direct strategy for multi-step modelling to help predict timesteps non-autoregressively and thus circumvents the error accumulation problem. The model was evaluated against state-of-the-art baselines using two benchmark datasets. Experimental results demonstrated the model's superiority, achieving up to 21.17% and 29.30% average improvements in Root Mean Squared Error and 3.56% and 32.80% improvements in Mean Absolute Error compared to the baselines, respectively. The Friedman Chi-Square statistical test further confirmed the significant performance difference between the proposed model and its counterparts. The adversarial perturbations and random sensor dropout tests demonstrated its good robustness. On top of that, it demonstrated good generalizability through extensive experiments. The model effectively mitigates error accumulation in multi-step predictions while maintaining computational efficiency, making it a promising solution for enhancing Intelligent Transportation Systems.