This study aims to analyze the dynamics of wildfires in the United States by predicting fire size for both small and large fires and developing a classification model for fire size. Using a dataset spanning from 1992 to 2015, machine learning models such as XGBoost, CatBoost, Random Forest, Generalized Linear Models (GLMs), and Mult-layer Perceptron (MLP) were applied to predict fire size, with XGBoost and CatBoost showing strong performance in predicting small and large fires, respectively. Additionally, classification models, including XGBoost, CatBoost, and Random Forest, were developed to distinguish between small and large fires, with challenges arising from class imbalance. Future work will focus on improving model performance by incorporating more detailed environmental data and exploring advanced machine learning techniques.

Many engineering, machine learning, and scientific issues require black-box optimization since direct analytical gradients are unavailable and objective evaluations are prohibitively expensive. Although existing evolutionary heuristics, Bayesian Optimization approaches, and Energy-Based Models (EBMs) have improved search efficiency, they frequently fail to retain scalability and stability in high-dimensional, dynamically changing landscapes. In this paper, we offer two novel frameworks, EBMBO and REBMBO, that use Bayesian uncertainty quantification, energy-based latent exploration, and reinforcement learning-driven adaptive decision-making to better navigate complex objective spaces. Our rigorous empirical examination reveals dramatically improved sample efficiency, faster convergence, and robust performance across a variety of optimization settings. This study pushes the frontier of black-box optimization by bringing together the capabilities of probabilistic modeling, structured exploration, and adaptive policy updates.

This research focuses on improving YOLOv8 for detecting road objects from a pedestrian’s viewpoint. It involves training three pre-trained models (YOLOv8n, YOLOv8s, YOLOv8m) on over 10,000 images, which include both a self-collected dataset of road objects and a subset from the COCO dataset. The study employs transfer learning to maintain the models’ proficiency in recognizing the original COCO dataset classes while integrating seven new categories. The models’ effectiveness was gauged using metrics such as precision, recall, mAP, and processing speed to identify the most suitable model for real-time road detection. Ultimately, the YOLOv8m model showed superior accuracy and reasonable processing speed, though its performance still falls short of real-world detection requirements.

This paper presents the idea of predicting Apple stock price using three different types of time series models based on many features including news headlines related to Apple. To collect data on if the news headlines have positive, neutral, or negative information, we used LSTM (Long Short Term Memory), GRU (Gated Recurrent Unit), BERT (Bidirectional Encoder Representations from Transformers) Sentimental Analysis, and BERT Fine-Tuning. The BERT Fine-Tuning model has the best result. For Apple stock price forecasting we used the classical time series models, ARIMA (Autoregressive Integrated Moving Average) and SARIMAX (Seasonal Autoregressive Integrated Moving Average with Exogenous Regressors), the supervised models, linear regression using PCA (Principal Component Analysis) and random forest, and the deep learning model, LSTM. The linear regression model using PCA performed the best. For further investigation, we could change the parameters of the LSTM model to get better results.

Earthquake is the most destructive hazard in building design; base-isolations, as one effective method mitigating the earthquake hazard, are widely used in the building design. However, simulating the building response under earthquake using the physical-based model is time-consuming and undesirable. Therefore, several statistical methods (linear regression, weighted least square, the decision tree, random forest, and neural network) are applied to predict building responses based on the characteristics of applied earthquakes. After principal component analysis, the Statistical models' prediction matches the simulation data very well, indicating that it is promising to utilize statistical methods in predicting the critical building response under earthquake. These predictions provide insightful guidance to the designer.

This thesis investigates the development of artificial intelligence systems with embodied understanding of the three-dimensional world. Moving beyond the limitations of current AI models that function as disembodied information processors, I present foundation models for machines that actively perceive, reason about, and interact with physical reality. The research progresses through four interconnected components that collectively bridge the gap between computational pattern recognition and embodied intelligence. ``The Thinking Eye" enables visual systems to reason about physics and causality, developing models that can go beyond sheer pattern recognition. "The World inside `I'" builds internal representations of 3D environments through novel descriptor field frameworks that allow machines to construct and update mental models of space through active exploration.In ``The Thinking Body: Reasoning and Acting in the 3D World," I introduce a set of 3D-based large language models that integrate multiple sensory modalities while actively interacting with 3D environments through perception-action loops. The final part, ``The Embodied Mind," examines how agents develop individual minds through accumulated experiences and how multiple agents with distinct histories can develop collaborative intelligence. Collectively, this research establishes a foundation for AI systems that understand the 3D world through grounded physical experience rather than pattern recognition.