This research focuses on the development of a system that employs genetic algorithms (GA) that generate increasingly challenging terrain for a player to navigate. Our system is developed within a video game, where a player is required to get from a start point to an endpoint, with gameplay mechanisms based on the terrain acting as obstacles for the player. The results from a game experience survey for the ingame section show that players rated gameplay more negatively when the terrain was randomly generated than when it was generated by the GA. While players did not feel more challenged using the GA-generated terrain, they gave an overall more positive gameplay experience. A quantitative data analysis of data collected during gameplay indicates an increase as the rounds progressed, in the number of deaths, the time taken, and levels skipped when using the GA. In contrast, the randomly generated terrain shows no significant difference.
In recent years, swarm robotics has proven to be a promising approach to develop collective behaviors that permit to extend the capabilities of a single robot in solving complex tasks. Here, the major challenge consists of designing individual behaviors that lead to the emergence of the desired collective behavior. In this regard, several automatic design methods have been proposed over the last decades. In this work, we present a hybrid approach that combines two of the main techniques of such design methods: multi-agent reinforcement learning and neuro-evolution. The idea is to evolve a pre-trained individual policy, in such a way that we can learn an effective collective behavior for vision-based navigation tasks. Specifically, we want a homogeneous swarm of low-cost robots to autonomously explore a simulated indoor corridor, while avoiding static and dynamic obstacles in a flexible and intelligent way. Through several experiments, we show how the learned collective policy outperforms the individual one in solving the previous task. Furthermore, our approach becomes highly memory-efficient, thanks primarily to a particular autoencoder able to compress input images into very compact visual representations.
Maintaining optimal blood glucose levels is paramount in diabetes management. Currently, patients are tasked with deciding when and how much insulin to administer. In contrast, reinforcement learning offers a closed-loop control system that automates this process. In this study, we present a reinforcement learning framework tailored for personalized blood glucose management, employing in silico patients. The agent dynamically adjusts insulin dosages in response to real-time glucose levels and patient-specific characteristics. We introduce HypoTreat, an innovative extension to the model that incorporates the 15-15 rule. This addition simulates the real-life patient’s behavior by temporarily opening the closed-loop, permitting virtual patients to consume food when their blood glucose level falls below 70, thus, preventing hypoglycemia.While the scenario of opening the closed-loop is undesirable, our findings reveal its efficacy in averting catastrophic hypoglycemia without increasing time spent in hyperglycemia. By introducing HypoTreat in training the average time spent in severe hypoglycemia has been reduced by 97.4%. Meanwhile, the in silico patients remain over 80% of their time within the target blood glucose range. This demonstrates the potential of our simulation-driven approach to enhance blood glucose control by intelligently incorporating real-life behavior, providing a safer and more effective solution for diabetes management.
The rapid increase in vehicular sensor data and advances in Internet of Things (IoT) technologies pose a dual challenge and opportunity for real-time traffic management and driver behavior analysis in the domain of Intelligent Transportation Systems (ITS). The current gap lies in effectively processing this data at the network edge, which is critical for timely and efficient decision-making in ITS. Our study proposes a novel, multi-layered, stream-oriented data processing methodology specifically designed for edge computing environments to address this challenge. This approach integrates soft sensors, the Typicality and Eccentricity Data Analytics (TEDA) framework, and an incremental clustering algorithm to detect and classify driver behavior patterns. The emphasis on using low-energy hardware and TinyML techniques is crucial, aiming to optimize processing efficiency while minimizing the environmental impact. To substantiate the efficacy of our methodology, we conducted a practical case study in Natal-RN, Brazil, utilizing the Freematics One + OBD-II microcontroller device for real-world application and validation. This involved two participants and focused on real-time data analysis for driver profile detection. The preliminary results demonstrate a significant potential of our approach in accurately classifying driving behaviors and patterns, offering insights for enhancing vehicle efficiency and reducing fuel consumption. This study fills a critical gap in ITS. It sets the stage for future research in sustainable and adaptive transportation systems, leveraging the power of edge computing and incremental algorithms in real-time data stream processing.
Large Language Models represent a transformative technology at the forefront of artificial intelligence and natural language processing, with applications spanning diverse domains. This study conducts a comprehensive science mapping analysis of the LLMs research field, leveraging bibliometric techniques to uncover its thematic structure, trends, and global actors involved. Utilizing data from the Web of Science, a corpus of 1303 research documents from 2010 to 2023 is analyzed, revealing a notable surge in research activity, particularly in recent years. Key thematic areas driving research within the field are identified, including chatbot, code generation, augmented reality, transformers, and machine learning paradigms. Foundational technologies such as transformers are pivotal in shaping the research landscape, while emerging themes like prompt learning hint at future directions. This study offers valuable insights for researchers, practitioners, and policymakers seeking to navigate the dynamic landscape of LLMs research and harness its full potential for societal benefit.
Temporal networks represent the evolving complex systems by regarding the contained elements as nodes and their connections as edges, respectively, which are both time-varying. Link prediction on temporal networks is an essential problem in real-world applications, which aims to forecast the evolution of temporal networks by predicting the future links to appear. However, existing methods generally focus on modeling the individual historical temporal features of source node and target node, while neglecting the complex correlations between them, thus leading to the suboptimal performance. In this paper, we propose a Correlation-enhanced Dynamic Graph learning (CoDyG) method to simultaneously take the individual features of source/target nodes and their correlations into consideration. Specifically, we achieve this by (1) introducing a co-attention network in the source/target node representation learning and (2) designing a temporal difference encoding strategy to model the temporal correlations between source/target nodes. Comprehensive experiments conducted on two widely adopted real-world temporal network datasets demonstrate that our proposed CoDyG can achieve the state-of-the-art performance in terms of the Average Precision (AP) and Area Under the Curve (AUC) metrics on the temporal link prediction task.
We introduce a modified incremental learning algorithm for evolving Granular Neural Network Classifiers (eGNN-C+). We use double-boundary hyper-boxes to represent granules, and customize the adaptation procedures to enhance the robustness of outer boxes for data coverage and noise suppression, while ensuring that inner boxes remain flexible to capture drifts. The classifier evolves from scratch, incorporates new classes on the fly, and performs local incremental feature weighting. As an application, we focus on the classification of emotion-related patterns within electroencephalogram (EEG) signals. Emotion recognition is crucial for enhancing the realism and interactivity of computer systems. The challenge lies exactly in developing high-performance algorithms capable of effectively managing individual differences and non-stationarities in physiological data without relying on subject-specific information. We extract features from the Fourier spectrum of EEG signals obtained from 28 individuals engaged in playing computer games – a public dataset. Each game elicits a different predominant emotion: boredom, calmness, horror, or joy. We analyze individual electrodes, time window lengths, and frequency bands to assess the accuracy and interpretability of resulting user-independent neural models. The findings indicate that both brain hemispheres assist classification, especially electrodes on the temporal (T8) and parietal (P7) areas, alongside contributions from frontal and occipital electrodes. While patterns may manifest in any band, the Alpha (8-13Hz), Delta (1-4Hz), and Theta (4-8Hz) bands, in this order, exhibited higher correspondence with the emotion classes. The eGNN-C+ demonstrates effectiveness in learning EEG data. It achieves an accuracy of 81.7% and a 0.002933 interpretability using 10-second time windows, even in face of a highly-stochastic time-varying 4-class classification problem.
Cardiovascular diseases count among the most critical and propagated diseases worldwide. Artificial Intelligence [AI] generates promising results across various medical domains and can enhance cardiovascular diagnosis and treatment. However, contextual knowledge is crucial for multiple design decisions to avoid pitfalls, and the true success of the intelligent application can only be measured in relation to its intended clinical setting. Machine learning [ML] models are evaluated based on performance metrics whose interpretation is influenced by data distribution and tuning objective. The present paper introduces a domain-embedded approach aiming towards a reliable performance evaluation of ML models throughout the complete lifecycle. Our findings are illustrated for multi-label ECG classification in an emergency setting and calculated on open-source Physionet data. Finally, the resulting procedure is embedded as a contribution to Quality Gate Metrics within our generic and customizable methodology based on Quality Gates towards certifiable AI in medicine.
While contextual data plays an important role for the outputs generated by AI models, it has not been fully considered when providing explanations about how and why those models generated such outputs. In the current paper we delve into previous research studies in order to provide an overview about the different uses of contextual data when enhancing the explanations for AI models. The inclusion and exclusion criteria are presented and a brief review of previous papers is captured, categorized by different ways to include contextual data into AI models. This paper pretends to foster research in leveraging the potential of contextual information for enhancing AI explainability.
In recent years, the development of intelligent robotic systems capable of detecting and tracking individuals using computer vision and artificial intelligence has gathered attention in the field of robot-assisted applications. Such intelligent and collaborative robots that operate in the vicinity of humans make significant strides in various public and private sectors. This workin-progress paper discusses a preliminary integrated robot system that utilizes a lightweight object detection and facial recognition techniques, namely OpenCV Haar Cascade, VGGFace and other navigation and tracking packages to enable a robot to search, identify and follow a person based on real-time streaming video and sonar sensor inputs. This study establishes a preliminary model to enable seamless robot and human cooperation in unknown environments.
In this paper, we propose a novel evolving clustering algorithm for streaming data entitled EdgeCluster. The proposed algorithm is resource efficient, making it suitable for use at edge devices with limited storage and computational capacity. The EdgeCluster is capable of modeling and monitoring a streaming data phenomenon and identifying outlying behavior. In parallel with the monitoring, the EdgeCluster algorithm dynamically maintains the set of clusters that models the phenomenon’s normal behavioral scenarios by taking newly arrived data into account and updating the clustering model accordingly. The EdgeCluster algorithm is evaluated and benchmarked to another resource-aware stream clustering algorithm, EvolveCluster, in two experimental data scenarios using synthetic and real-world datasets.
Superpixel-based segmentation is an important pre-processing step for the simplification of image processing. The subjective nature behind the determination of optimal cluster numbers in segmentation algorithms can result in either underor over-segmentation burdens, depending on the image type. Insect wings, with their intricate color patterns, pose significant challenges for the accurate capture of color diversity in clustering algorithms, assuming a spherical and isotropic cluster distribution is used. This paper introduces a hybrid approach for color clustering in insect wings, integrating the Simple Linear Iterative Clustering (SLIC) method to generate the initial superpixels, and a DeltaE 2000 function the precisely discriminated merging of superpixels. Color differences between superpixels serve to measure homogeneity during the merging process. The proposed new algorithm demonstrates enhanced segmentation as it overcomes the issue of over-segmentation and under-segmentation, as evidenced by the results derived from the Boundary Recall, Rand index, Under-segmentation Error, and Bhattacharyya distance using ground truth data. The Silhouette score and Dunn Index are also used to quantitatively evaluate the efficacy of our new proposed clustering technique.
In a permanently connected society where everyone can share and discuss information of every kind, recognizing credible news or news sources has been demonstrated to be hard. Recent research is going toward the definition of tools that try to automate cognitive processes behind the evaluation of content trustworthiness and simplify people’s lives. However, despite recent efforts to identify fake or misleading content, analyzing sources of information in their entirety remains marginal or only a manual task. This paper aims to define a systematic approach for scoring the credibility of web domains. Specifically, it assesses the Multifactorial Source Credibility Score (MSCS) that, through collecting domain pages as a stream, estimates multiple factors such as the quality of content, the author’s trustworthiness, the presence of biases and advertisements, the source traffic, etc. Then, it evaluates a weighted moving average of web page scores to attribute an overall score to the domain, taking into account scores over time. Through experiments conducted on a real dataset, the study presents promising results by showcasing a notable correlation between the proposed indicator and a manually assigned score by NewsGuard.
This paper explores the nuances of extracting knowledge for auction verification, utilizing advanced methodologies from machine learning and fuzzy neural networks. In a meticulous comparison with state-of-the-art approaches, the proposed model demonstrates an improvement in accurately distinguishing diverse auction scenarios. The study not only underscores the effectiveness of these sophisticated technologies in enhancing auction verification processes but also underscores the crucial role played by knowledge extraction through fuzzy rules. This extraction process emerges as a valuable asset, particularly for the detection and mitigation of potential fraud in auction settings.
Group decision-making, an everyday practice ranging from simple choices to long-term investments, is influenced by the inherent ambiguity of the natural language we use as people. The ambiguity in this instance stands in stark contrast to the precision of language employed by machines. Furthermore, expressing emotions, such as aggressiveness or happiness, in natural language can pose a problem by influencing the group decision-making process. To address this challenge, we propose a consensus method for group decision-making that, through sentiment analysis, assesses aggressiveness in comments using generative artificial intelligence, specifically the large language models provided by Google. Using a Gemini API, we compute the degree of hostility in the comments, adjusting the weight of each expert. This ensures that more aggressive experts have less influence on the decision-making process. In such a way, this tool is able not only to encourage more objective language but also to prevent experts from unduly influencing others through their comments.
Renewable energy offers a sustainable solution to climate change and environmental degradation by reducing greenhouse gas emissions and pollution. Solar energy, in particular, is a promising source due to its global availability, environmentally friendly operation, and declining costs. However, faults in solar photovoltaic (PV) systems can diminish their efficiency and pose safety risks. Detecting and diagnosing these faults is crucial for optimizing performance, minimizing downtime, and ensuring safety in solar PV installations. Contextual data, widely used in fields like NLP, computer vision, and recommender systems, are not as prevalent in the domain of solar PV systems. This paper focuses on leveraging contextual information to enhance fault detection in PV systems using ML-based models. The objective is to explore methods for integrating contextual features into machine learning models for AC power prediction and fault detection in solar PV systems. Three strategies for handling contextual data are compared. The results showed that for AC power prediction, contextual expansion and contextual normalisation performed better than contextual model selection strategy.
In this paper, we introduce an enhanced Particle Swarm Optimization using reflection strategy called Reflected Particle Swarm Optimization (RPSO) algorithm, RPSO designed to enhance robot path planning in complex environments such as static environment with obstacles. The integration aims to improve the exploration and exploitation capabilities of the search process. RPSO has been compared with state-of-art and well-known optimization algorithms. The results demonstrate that RPSO consistently outperforms the compared algorithms in finding the shortest paths and maintaining high levels of consistency across different trials. Furthermore, the results show that RPSO, with its innovative reflection mechanism, offers a significant advancement in robot path planning.
Vehicle platooning is a promising solution to improve traffic behaviour for autonomous vehicles. The advantages of this mode of transportation include decreased travel time, reduced pollution, improved fuel consumption, and the potential to avoid traffic congestion. However, the technology is not yet mature enough due to several technical challenges that complicate its development in real systems, such as vehicle heterogeneity, the presence of disturbances such as wind on vehicle longitudinal dynamics, and delays and interruptions in communication networks, among others. This paper examines the individual effects of each aspect to determine their order of relevance and establish a framework for designing robust vehicle platoon controllers to handle such effects.
This paper presents a novel algorithm that integrates computer vision and deep learning to address the critical task of predicting the short-term trajectory of beet armyworm larvae. The algorithm integrates techniques such as object classification, segmentation, tracking, image processing, and time-series predictive modeling. The algorithm utilizes You Only Look Once fifth generation (YOLOv5) with a mAP score of 96.27% and ByteTrack for larval tracking, generating five consecutive frames of larval contours and trajectories. A specialized image processing algorithm extracts significant features for the chosen larva in only 0.011 seconds per frame. Subsequently, the Long Short-Term Memory (LSTM) model accurately predicts the trajectories of larvae. The average Mean Absolute Error (MAE) of 0.2573 millimeters for motionless larvae and 0.7444 millimeters for crawling larvae demonstrates its precision. The algorithm is capable of predicting the central point locomotion of larvae 1.5 seconds ahead. This algorithm serves as a foundational reference for implementing intelligent automation in the field of insect rearing.
Long Short-term Cognitive Networks (LSTCNs) are recurrent neural networks for univariate and multivariate time series forecasting. This interpretable neural system is rooted in cognitive mapping formalism in the sense that both neural concepts and weights have a precise meaning for the problem being modeled. However, its weights are not constrained to any specific interval, therefore conferring to the model improved approximation capabilities. Originally designed for handling very long time series, the model’s performance remains unexplored when it comes to shorter time series that often describe real-world applications. In this paper, we conduct an empirical study to assess both the efficacy and efficiency of the LSTCN model using 25 time series datasets and different prediction horizons. The numerical simulations have concluded that after performing hyper-parameter tuning, LSTCNs are as powerful as state-of-the-art deep learning algorithms, such as the Long Short-term Memory and the Gated Recurrent Unit, in terms of forecasting error. However, in terms of training time, the LSTCN model largely outperforms the remaining recurrent neural networks, thus emerging as the winner in our study.
