Abstract
The gold standard real-time core temperature (CT) monitoring methods are invasive and cost-inefficient. The application of the Kalman filter for an indirect estimation of CT has been explored in the literature for more than 10 years. This paper presents a comparative study between different state-of-the-art Extended Kalman Filter (EKF) approaches. Moreover, we proposed the addition of an extra layer to the pipeline that applies a pre-emptive mapping concept based on the physiological response of the heart rate (HR) signal, before using it as input to the EKF. The algorithm was trained and tested using two datasets (18 subjects). The best-performing approach with the novel pre-emptive mapping achieved an average Root Mean Squared Error (RMSE) of 0.34 degrees C, while without pre-emptive mapping, it resulted in an RMSE of 0.41 degrees C, leading to a performance improvement of 17%. Given these favorable outcomes, it is compelling to assess the efficacy of this method on a larger dataset in the future.

Abstract
Time series data are essential for a wide range of applications, particularly in developing robust machine learning models. However, access to high-quality datasets is often limited due to privacy concerns, acquisition costs, and labeling challenges. Synthetic time series generation has emerged as a promising solution to address these constraints. In this work, we present a framework for generating synthetic time series by leveraging complex networks mappings. Specifically, we investigate whether time series transformed into Quantile Graphs (QG) -- and then reconstructed via inverse mapping -- can produce synthetic data that preserve the statistical and structural properties of the original. We evaluate the fidelity and utility of the generated data using both simulated and real-world datasets, and compare our approach against state-of-the-art Generative Adversarial Network (GAN) methods. Results indicate that our quantile graph-based methodology offers a competitive and interpretable alternative for synthetic time series generation.

Abstract
Efficient task allocation and coordination are critical for heterogeneous multi-agent systems operating in dynamic field environments. This paper presents a closed-loop framework that integrates Large Language Models (LLMs) with graph-based optimisation to enable end-to-end task decomposition, allocation, and adaptive execution. High-level task scripts are initially parsed by an LLM into structured execution flows, which are transformed into Directed Acyclic Graphs (DAGs) capturing action-level dependencies. A Genetic Algorithm (GA) then optimises agent-to-task assignments by minimising makespan under capability and battery constraints. To ensure robustness, the framework incorporates an LLM-driven recovery module that enables localised replanning under execution failures without interrupting unaffected agents. System-level experiments in a high-fidelity agroforestry simulation demonstrate a 37% increase (p<0.001) in harvesting productivity and a 19% reduction in human idle time compared to manual baselines. Under mid-execution failures, the system maintains significantly higher performance, with replanning latencies averaging 24 s. The framework scales to large fleets (up to 1000 agents) and effectively enhances human-robot collaboration through structured, dependency-aware coordination.

Abstract
Hexapod robots, with their six-legged design, excel in stability and adaptability on challenging terrain but pose significant control challenges due to their high degrees of freedom. While reinforcement learning (RL) has been explored for robot navigation, few studies have systematically compared on-policy and off-policy methods for multi-legged locomotion. This work presents a comparative study of SARSA and Q-Learning for trajectory control of a simulated hexapod robot, focusing on the influence of learning rate (alpha), discount factor (gamma), and eligibility trace (lambda). The evaluation spans eight initial poses, with performance measured through lateral deviation (Ey), orientation error (E theta), and iteration count. Results show that Q-Learning generally achieves faster convergence and greater stability, particularly with higher gamma and lambda values, while SARSA can achieve competitive accuracy with careful parameter tuning. The findings demonstrate that eligibility traces substantially improve learning precision and provide practical guidelines for robust RL-based control in multi-legged robotic systems.

Abstract

Abstract
This study presents a methodology for forecasting municipal water consumption to support efficient resource management. Using monthly data from 2018 to 2022 for the municipality of Vouzela, Portugal, three forecasting approaches were evaluated: SARIMA, Holt-Winters, and TimeGPT. Data preparation included logarithmic transformation and stationarity assessment using the KPSS test, ensuring appropriate conditions for statistical modelling. The SARIMA model was selected automatically based on the Akaike Information Criterion (AIC), while the Holt-Winters method was fitted with additive components and a Box-Cox transformation. In addition, TimeGPT was employed as a state-of-the-art foundation model for time series forecasting. The three methods were used to predict water consumption for the 12 months of 2023, and their performance was assessed using MAE, MSE, RMSE and MAPE. Results indicate that although all methods perform adequately, Holt-Winters and TimeGPT better capture recent consumption dynamics, providing more accurate forecasts in several periods. Overall, this study shows that combining classical statistical models with advanced forecasting techniques offers local authorities reliable and computationally accessible tools to support water supply planning and sustainability.