This study focuses on the sea lampreys that have invaded the Great Lakes. Considering the significant impact of their sex ratio changes with the environment on the ecosystem, the aim is to construct a model to explore its mechanism. Data from various sources such as government departments, scientific research institutions, and academic articles were collected and preprocessed. The Lake Ontario Lamprey Ecosystem Dynamics Model (OLED) was created, which consists of the Sex Ratio Relationship (SRR), Lamprey Life Cycle (LLC), and Lake Ontario Species Survival (LOSS) modules. The Lasso regression and other algorithms were used to analyze the impact of environmental factors on the sex ratio. Based on the Lotka-Volterra and other models, the relationships between sea lampreys and other species were studied. The Gray Forecast Model was also adopted to predict relevant variables and calculate the sex ratio. The results show that the changes in the sex ratio of sea lampreys significantly affect the ecosystem, such as influencing fish populations and nutrient cycling. Moreover, the model is robust under small data perturbations and is sensitive to the reproduction rate. This research provides theoretical support for the ecological control of sea lampreys.

This research proposes a narrative episodic memory framework for generative AI-based NPCs in games, structured around the "Who, What, When, Where, Why" (5Ws) format, which aims to suggest a memory structure for future GenAI NPC model development. The main objective is to allow NPCs to recognize and remember memories of interactions with players and to respond in a way that is contextually rich and consistent with their personality and background. The framework implements the idea of memory decay, enabling NPCs to tune their personality and tone for different players using generative AI over time. By analyzing the interaction memories, NPCs can dynamically adjust their character to create a personalized and immersive experience. The framework will be evaluated using several large language models, and their responses will be compared through analysis. The results will be presented with an emphasis on the differences among models to cater to the distinct needs of developers. While such a framework is considered a component that needs integration, this work establishes a foundation for designing episodic memory structure, allowing NPCs to participate with players with complex memory structure as it allows them to engage in more meaningful interactions. This foundational framework serves as pathways for future development on NPC AI that enables sophisticated interaction and changes character personality specific to the way players engage with them.

Human body recognition is crucial for enhancing security, facilitating human-robot interaction, and improving accessibility for people with disabilities. The integration of deep learning techniques has revolutionized the field, significantly boosting the accuracy and efficiency of body recognition systems. This advancement not only improves security but also enriches the user experience in various applications, from healthcare to entertainment. This study explores the application of Convolutional Neural Networks (CNNs), Vision Transformers (ViTs), Selective Kernel Networks (SKNs), and Adaptive Kernel Convolution (AKConv) in identifying individuals from a distance. Leveraging transfer learning from large-scale datasets like ImageNet, evaluation of these models on a standardized human body recognition dataset, focusing on the trade-off between recognition performance and computational efficiency. Findings underscore the potential of SKNs and AKConv in achieving high accuracy with reduced computational demands, paving the way for their deployment in resource-constrained environments. The research contributes to the development of more efficient recognition algorithms and provides insights for future advancements in the field.

Point cloud completion is an important problem in 3D computer vision that aims to reconstruct the missing or occluded parts of point cloud data, often captured through LIDAR or depth sensors. Traditional methods have struggled to effectively complete these point clouds while maintaining fine details and preserving global context, particularly in the presence of complex occlusions and sparse data. In this paper, we introduce a novel deep learning architecture for point cloud completion, named PC-GraphNet. This method combines the strengths of three powerful paradigms: point-based networks, graph convolutional networks (GCNs), and generative adversarial networks (GANs). PC-GraphNet leverages a hybrid approach that integrates a point-wise attention mechanism for local feature refinement, a GCN for global contextual reasoning, and a GAN-based decoder for generating realistic completions. By incorporating both local and global feature extraction, our model significantly improves the quality of completed point clouds, especially in cases of occlusions and sparse input data. Our experiments demonstrate that PC-GraphNet outperforms state-of-the-art methods in terms of both completion accuracy and visual fidelity. Additionally, the model is computationally efficient, making it suitable for real-time applications.

With the acceleration of population aging, the demand for elderly care services is growing rapidly, and the shortage of caregivers has become a global social issue. To address this challenge, the application of artificial intelligence technology in the field of elderly care, and the development of assistive robots capable of providing personalized and efficient care services, has become a focus of research. This study aims to develop an intelligent robot that can understand elderly people's instructions, perceive the surrounding environment, and provide personalized assistance services by integrating advanced artificial intelligence technologies such as image recognition and natural language processing into the robotic system. Specifically, the robot will have the following functions: First, the robot can conduct natural language interaction. Through speech recognition and semantic understanding technology, it can accurately recognize the elderly person's voice commands and conduct natural and fluent conversations. Second, the robot can realize visual perception. By utilizing image recognition technology, it can perceive the surrounding environment in real-time, recognize objects, faces, etc., and respond accordingly based on the scene. Finally, based on the understanding of instructions and the perception of the environment, the robot can complete a series of daily tasks such as fetching objects, companionship, and reminders. By developing and evaluating such intelligent robots, we hope to improve the quality of life for the elderly, reduce the workload of caregivers, promote innovation in elderly care service models, and provide technical support and theoretical foundation for future intelligent elderly care.

The paper presents how ABM, RL and NNs are combined to simulate the behaviour of investors in financial markets. ABM is an adaptable approach to modeling highly complex financial systems in which agents (investors) act and change on incomplete or distorted data. Reinforcement Learning optimizes agent behaviour through learning from their mistakes and making changes over time Neural Networks optimise the model by finding non-linearity in the market data to enhance the model’s predictive ability. The research in this paper examines how misguided investor decisions, such as overconfidence and herd behavior, cause market crashes, bubbles and crashes, even when agents have taken some lessons from their experiences. The results show that ABM, RL and NNs work together in way that provide valuable insights into the nature of financial markets, and reveal the promise and limitations of applying AI to predict complex and unpredictable markets. These results indicate that although Neural Networks predict more accurately when markets are stable, they are poorly effective when market is unstable with very high irrationality. These insights may be used to improve the regulation and policymaking of markets in practice.

This paper introduces VisionSweep, a novel smart cleaning system designed to address current limitations in household robotic cleaning by leveraging advanced vision and path-planning technologies. Traditional cleaning robots often lack adaptability, resulting in inefficient cleaning patterns and unnecessary energy consumption. VisionSweep overcomes these challenges through a camera-based vision system that captures human movement patterns and identifies areas requiring cleaning. By employing image preprocessing and the Single Shot Detector (SSD) model, the system accurately detects dirt and high-traffic zones. A path-planning algorithm, such as Dijkstra’s, optimizes the cleaning route based on detected dirt and human activity, ensuring that the cleaning robot prioritizes key areas. This solution offers a promising approach to resource-efficient and effective automated cleaning within smart home environments.

The rising prevalence of mental health issues identifies the urgent need for accurate, scalable, and timely prediction systems. Deep learning, a subset of a machine learning inspired by human’s neuron structure, has offered an opportunity for innovative solutions for mental health diagnosis. The main idea of this paper is analyzing the application of deep learning in diagnosing mental disorders, including but not limited to Alzheimer, Parkinson and Schizophrenia. An enormous number of techniques will be put into real life application while dealing with diagnosis of mental health issues, but our focus will be convolutional neural networks (CNN), recurrent neural networks (RNN) and deep neural networks (DNN) strategies this time. On top of that, we’ll also discuss integrating multimodal data, combining neuroimaging with behavioral and textual analysis which is able to enhance diagnostic accuracy. Techniques like CNN-DNN hybrids provide another comprehensive view of mental health issues. Despite these successes, challenges such as data privacy, ethical considerations, and security issues remain significant obstacles, which is our topic as well. Future prospects include federated learning, blockchain technology, and privacy-preserving methods to enhance security and maintain patient trust. All in all, this paper aims at giving a comprehensive understanding of deep learning methods through the application of the mental health diagnosis area.

Semi-Supervised Learning (SSL) is implemented when algorithms are trained on both labeled and unlabeled data. This is a very common application of ML as it is unrealistic to obtain a fully labeled dataset. Researchers have tackled three main issues: missing at random (MAR), missing completely at random (MCAR), and missing not at random (MNAR). The MNAR problem is the most challenging of the three as one cannot safely assume that all class distributions are equal. Existing methods, including Class-Aware Imputation (CAI) and Class-Aware Propensity (CAP), mostly overlook the non-randomness in the unlabeled data. This paper proposes two new methods of combining multiple imputation models to achieve higher accuracy and less bias. 1) We use multiple imputation models, create confidence intervals, and apply a threshold to ignore pseudo-labels with low confidence. 2) Our new method, SSL with De-biased Imputations (SSL-DI), aims to reduce bias by filtering out inaccurate data and finding a subset that is accurate and reliable. This subset of the larger dataset could be imputed into another SSL model, which will be less biased. The proposed models have been shown to be effective in both MCAR and MNAR situations, and experimental results show that our methodology outperforms existing methods in terms of classification accuracy and reducing bias.

The study presents a new optimization method that merges Backpropagation Neural Networks (BPNN) with an enhanced Energy Valley Optimizer (EVO) algorithm to elevate the performance of multi-input nonlinear systems. Traditional optimization methods like Gradient Descent and Particle Swarm Optimization (PSO) face issues with slow convergence and local minima entrapment which the hybrid model successfully resolves. The exploitation of EVO's innate energy dissipation mechanism results in faster BPNN convergence and improved generalization capabilities. Experiments with both synthetic and real-world datasets proved that the hybrid model showed better accuracy and prediction capabilities along with faster convergence time. The findings demonstrate substantial decreases in Mean Squared Error (MSE) and Root Mean Squared Error (RMSE) relative to conventional optimization methods. The sensitivity analysis demonstrates that optimal EVO performance depends on precise adjustments of parameters including mutation rate and local search operator. This study presents an efficient method to enhance BPNN training which shows promise in managing complex nonlinear systems.