32584 research outputs found
Sort by
BlindFun: Transforming a Video Game to Become Blind Accessible
No full textMany video games are not fully playable to visually impaired players due to prioritizing visual fidelity over gameplay. This study goes over the development of our 2D shooter, specifically designed for blind players. The game aims to balance challenge and accessibility, where other accessible games tends to prioritize one over the other. The game integrates spatial audio for combat and navigation mechanics, situational sound cues, and text-to-speech (TTS) messages to convey game state information. This study used an iterative design approach to develop two game builds, each playtested across three sessions with visually impaired and sighted participants. The first test focused on game flow and identified accessibility issues, such as difficulty collecting power-ups and gauging enemy distance. To solve these problems, we implemented condition-based item spawning and Game-AI techniques, like FSM driven enemy behavior to enhance gameplay. The second session focused on engagement, improvements from the previous session. The final playtest collected qualitative and quantitative feedback comparing accessibility features across builds. Results from both sessions showed improved player gameplay, addressing better accessibility. Ratings showed that refinements in build 2 enhanced features like enemy proximity and item collection. These results emphasize the importance of inclusive game design to improve accessibility
Optimizing Analytics Storage Strategies for Search Engines and Wiki Platforms
No full textModern search engines and wiki platforms generate vast quantities of user inter-/= / \u3eaction data such as page views, edits, clicks, and session events. This data must be stored and aggregated efficiently to enable scalable analytics and responsive querying. Yioop, an open-source search engine framework, serves as our primary case study, processing millions of such events to power its indexing and recommendation features. This report explores a shift from the conventional database storage based model to a log based model, in order to improve scalability and write efficiency. A size-limited, append-only logging facility was provided to log analytics events: the active log file rotates when the defined size limit is reached, keeping up file sizes manageable. Existing aggregate routines were modified to roll up logged events into summary records for efficient querying. The paper discusses performance benchmarks comparing the two storage models on storage footprint, write latency, response time of the query, and overall throughput under differing loads. Results demonstrate that the log-based method reduces storage footprint significantly and scales well with increasing data, delivering higher write throughput in sequential ingestion tests and keeping pace with the database under concurrent load, while maintaining comparable read performance. These findings confirm the log-based approach as a good way forward for Yioop’s analytics infrastructure and identify opportunities for additional optimization of data compaction and indexing
Domain-Specific Graph RAG Pipelines: Optimized Approaches for Building Efficient Personal Knowledge Repositories
No full textManaging personal data, including notes, calendar events, to-do lists, and other personal information, has become increasingly complex and challenging. In response to this issue, I propose a framework using RAG that enables a large language model (LLM) to efficiently query this data without requiring training on the personal data itself. Conventional retrieval systems, including those leveraging vector-based retrievalaugmented generation (RAG), are effective at handling basic queries but struggle to deliver coherent global abstractions, integrate diverse knowledge sources, and account for temporal nuances. This research explores domain-specific Graph based RAG frameworks that incorporate a knowledge graph to better model relationships, thereby enabling more comprehensive reasoning. By optimizing graph construction and modeling temporal entities for enhanced understanding, this research aims to advance the capabilities of RAG systems beyond the limitations of vector-based approaches for personal knowledge management
Social Engineering Scenario Generation for Awareness-Based Attack Resilience
No full textSocial engineering is found in a strong majority of cyberattacks today, as it is a powerful manipulation tactic that does not require the technical skills of hacking. Calculated social engineers utilize simple communication to deceive and exploit their victims, all by capitalizing on the vulnerabilities of human nature: trust and fear. When successful, this inconspicuous technique can lead to millions of dollars in losses. Social engineering is not a one-dimensional technique; criminals often leverage a combination of strategies to craft a robust yet subtle attack. In addition, offenders are continually evolving their methods in efforts to surpass preventive measures. A common utility to defend against social engineering attacks is detection-based software. Security awareness, however, is a valuable approach that is often eclipsed by automated tech solutions. Awareness establishes a strong first line of defense against these ever-changing attacks. This study utilizes three data-supplemented large language models to generate custom social engineering scenarios with the goal of supporting strong example-driven security awareness programs. The performances of BERT, GPT-3.5, and Llama 3.1 are comparatively analyzed, with Llama 3.1 producing the highest quality scenarios based on a series of metrics, including LLM-as-a-judge
AI-Enabled Anticipatory Handover Predictions in 5G Networks
No full textAs users move across network cells in 5G, it is critical to maintain seamless connectivity through efficient and fast handovers. However, as 5G networks have a very dense deployment of cells and higher carrier frequencies, handovers are often more frequent and challenging, leading to failures or the ping-pong effect. In this research, we are going to use Artificial Intelligence (AI) techniques to enable predicting handover(HO) events proactively as opposed to reactively, aiming to reduce HO failures and unnecessary handovers. We develop Long-Short-Term Memory (LSTM) and Bidirectional LSTM (Bi-LSTM) models to forecast future signal measurements, predict handover trigger points, and compare them with classical approaches like Random Forest(RF) and K-Nearest Neighbors(KNN). In our proposed approach, we also use oversampling methods such as the synthetic minority oversampling technique (SMOTE) and temporal SMOTE (T-SMOTE) to address class imbalance in HO event data. The implementation results demonstrate that deep learning models (LSTM and Bi-LSTM) achieve higher accuracy and F1 scores of rare handover events compared to KNN and RF. In addition, the Bi-LSTM model slightly outperforms the unidirectional LSTM, highlighting the benefit of capturing temporal patterns in both directions. These findings suggest that AI-driven handover prediction can significantly improve preparedness to receive handovers in 5G networks
Suicidal Ideation Detection on Reddit using LLM-Annotated Data and Graph Neural Networks
No full textSuicide is the fourth leading cause of death among people aged 15-29. More than 720, 000 people commit suicide every year. During the COVID-19 pandemic, we saw an increase in people seeking out mental health support on anonymous forums like Reddit. These anonymous forums allow people to express their suicidal ideation without judgment and give them a support structure that not everyone has. The aim of this project is to detect suicidal ideation using Reddit. In this work, we propose SIRGEL (Suicidal Ideation on Reddit using Graph Embeddings and LLMs), a dual-pipeline approach that combines large language model (LLM)- based annotation with both traditional text classification and graph neural network (GNN)-based modeling. The first step is to create a Reddit dataset for related subreddits. The second step is to annotate them as suicidal or not suicidal using LLMs. We also detail the creation of a new dataset made using Reddit posts and its benchmarking using baseline machine learning models. We model this problem using two approaches: traditional text classification and graph representation learning. Additionally, experiments with different types of Graph Neural Networks are performed with optimizations to improve and achieve efficient result
Addressing Urban Traffic Congestion: A Deep Reinforcement Learning-Based Approach
Get PDFIn an innovative venture, the research team embarked on a mission to redefine urban traffic flow by introducing an automated way to manage traffic light timings. This project integrates two critical technologies, Deep Q-Networks (DQN) and Auto-encoders, into reinforcement learning, with the goal of making traffic smoother and reducing the all-too-common road congestion in simulated city environments. Deep Q-Networks (DQN) are a form of reinforcement learning algorithms that learns the best actions to take in various situations through trial and error. Auto-encoders, on the other hand, are tools that help simplify complex data, making it easier for the DQN to understand and make decisions. To enhance the accuracy of these decisions, the research team chose average vehicle speed as a crucial indicator of traffic flow and employed HyperOPT, a method for fine-tuning the system’s hyper-parameters. The team put their method to the test in three different traffic scenarios: controlling a single intersection, managing multiple intersections, and overseeing protected left-turn signals. The results were clear and promising. The innovative system significantly improved traffic conditions by either reducing the average wait time atlights or increasing the overall speed of vehicles passing through intersections. This research not only presents a leap forward in traffic management but also offers a glimpse into a future where road congestion could be significantly alleviated. By employing cutting-edge AI and data processing techniques, the project stands as a testament to the potential for smart cities where traffic flow is optimized, making commutes faster and safer for everyone
An Ambient NOx Reduction Reactor System with and without Forced Air
Get PDFAmbient nitrogen oxide (NOX) is one of the main ingredients involved in the formation of ground-level ozone (smog), which contributes to global warming and the formation of particulates in the atmosphere and causes respiratory illnesses, especially in children, older adults, and people with lung diseases. Investigations on adding titanium dioxide (TiO2) nanoparticles to the paint used on concrete surfaces such as buildings and its potential in reducing ambient NOX concentration have provided opportunities for the research and development of methods for passive control and filtration of the ambient NOX. The present research used numerical geometrical optimization to identify an optimized surface geometry of half-stepped cylinders for increasing air recirculation and residence time for maximum ambient NOX reduction. The optimized surface was painted with titanium-infused paint and tested for NOX reduction according to ISO 22197-1 standard. Results showed a maximum of 11.5% increase in NO2 formation for an optimized surface compared with the result for the untextured smooth surface, with an estimated NOX reduction of approximately 15%. The results provided guidelines for developing a catalyst system for reducing ambient NOX when UV light is present. These results demonstrate potential of using paint infused with titanium to help reduce NOX and thus reduce smog, which harms people and the environment
Application of Root Cause Analysis for Fall Prevention: A Quality Improvement Initiative for Older Adults in a Skilled Nursing and Long-Term Care Facility
Get PDFFalls in older adults are common and often have severe outcomes. They are the leading cause of fatal and non-fatal injuries among people aged 65 and older and continue to increase. Half of residents in nursing facilities fall annually, and one in every ten falls will lead to a severe injury. Federal regulations like the Centers for Medicare and Medicaid Services require long-term care (LTC) facilities to incorporate quality assurance performance improvement initiatives to address identified quality concerns in LTC facilities. Root cause analysis (RCA) can help clinicians identify several root causes of falls and guide clinicians in developing personalized approaches to preventing falls. This quality improvement (QI) project sought to implement the RCA process to reduce fall rates among skilled nursing and LTC residents aged 65 and older. A quasi-experimental design was employed, and fall rates were compared three months before and three months after the implementation of RCA. Results showed a reduction of 3.14 falls per 1000 bed days and no fall-related fractures three months after the implementation of RCA. The findings revealed a marginally significant difference in fall rates, and residents’ history of falls was a significant predictor of falls. Utilizing the RCA process decreases the incidence of falls among older adults. It assists clinicians in identifying intrinsic, extrinsic, and environmental root causes of falls, helping them to develop multifactorial fall prevention strategies
Perceptions of Perinatal Depression Among Immigrant Mothers: Screening Practices in Nurse Home Visitation
Get PDFScreening for perinatal depression is a recommended strategy aimed at prevention and timely management of mood disruption in pregnancy and the postpartum (ACOG, 2024). Screening for perinatal depression is often integrated in evidence-based home visitation curriculums. The aim of this qualitative descriptive quality improvement project was to explore perceptions of perinatal depression among immigrant mothers engaged in nurse home visitation to identify barriers, if any, to disclosure and treatment of perinatal depression among a group at increased risk of developing perinatal depression (Gennaro, S., et al, 2016). Immigrant mothers residing in a large urban county in the San Francisco Bay Area, receiving perinatal care and engaged in an evidence-based, nurse led home visitation program were recruited. Data was collected from eligible participants through individual semi-structured interviews and analyzed utilizing thematic analysis, revealing four main themes (1) suffering (in silence) (2) isolation (3) fear and (4) lack of support. These findings revealed that, despite receiving perinatal care and engaged with regular nurse-led home visitation services, immigrant mothers are depressed in the perinatal period either because they are not screened or are fearful in disclosing their depression and mental health needs when they are screened. Perinatal depression among immigrant mothers engaged in home visitation is thus impacted at a critical stage, disclosure of symptoms, impacting the goals of screening practice recommendations and targeting of mood disorders in pregnancy and the postpartum. Nurses in home visitation are encouraged to anticipate and consider factors that impact depression screening responses and barriers to disclosure among their patients, especially immigrant mothers, to promote optimal perinatal mental health outcomes