Amber K Gupta

Abstract— This paper covers an approach to develop an RL agent
for the gym’s Lunar Lander problem, implementation of the RL
agent along with training and testing of the agent. Observations
during experiment runs are noted and results are analyzed.
Hyperparameter tuning and its impact on overall training is
presented in context of the environment and RL agent behavior.
Some future work is suggested based on the exploration of the
problem.
Keywords— Reinforcement Learning, MDP, Lunar Lander, DQN

Abstract— This paper covers observations and analysis during
and after running machine learning experiments on two different
public datasets. Observations by applying unsupervised machine
learning clustering techniques and Dimensionality reduction
techniques on the chosen dataset, are compared and contrasted.
The resulting clusters and dimensions are applied in the
supervised setting of the Neural Network for one of the datasets to
evaluate the effect of these data summarization techniques on
supervised learning. Effects of dimensionality reductions are also
observed on Unsupervised clustering methods.
Keywords— Machine Learning, unsupervised, Clustering,
Dimensionality Reduction, Breast Cancer, Waveform, KMeans,
EM, PCA, ICA, Random Projections, Random Forests

Abstract— This paper covers observations and analysis during
and after running random optimizer experiments in different
settings of optimization problems. Observations are compared and
contrasted to showcase the strength of different optimizers.
Keywords— Machine Learning, Random Optimizer, Compare
Models, Random Hill climbing, Simulated Annealing, Genetic
Algorithm and MIMIC, Breast Cancer

Here is abstract from the paper: more details on request…

MAESTRA (Multi-Agent Ecosystem of Scalable Testing, Review, and Assessment using LLMs in Cloud Configurations) is a novel multi-agent generic framework for automated cloud configuration validation that leverages Large Language Models (LLMs) to address the challenges of DevSecOps in cloud environments. Unlike single-LLM approaches, MAESTRA employs multiple specialized agents, each with specific roles (e.g., configuration validation, documentation retrieval, result verification, vulnerability analysis, and report generation), which work together. This collaborative approach overcomes limitations of single LLMs such as static knowledge cutoffs, hallucination risks, and the lack of real-time information, improving accuracy and reliability. MAESTRA integrates with existing DevSecOps workflows, enabling early detection of security vulnerabilities and errors. Evaluations across multiple LLMs and datasets, including a new real-world Terraform dataset, demonstrate improved performance compared to a single-LLM baseline, showcasing MAESTRA’s potential for robust and scalable configuration validation in modern cloud environments.

Abstract— This paper covers an approach to develop an
Multi-Agent RL (MARL) agent for a reduced scale Google’s
Football problem, implementation of the MARL agent along with
training and testing the agent to beat provided 3 baseline teams,
playing with Game AI. Key hypotheses to beat the baselines and
observations during experiment runs are noted and results are
analyzed. Some future work is suggested based on the exploration
of the problem.
Keywords— MARL, MDP, Football, PPO, Central Critic, QMIX

The report focuses on reproducing the results of a research paper titled "Real-world Patient Trajectory Prediction from Clinical Notes Using Artificial Neural Networks and UMLS-Based Extraction of Concepts."
Reproduction Implementation:
Demo Notebook: A Google Colab notebook was created to reproduce the results. The demo notebook includes comprehensive coverage of the comments, figures, and equations relevant to the project.
Presentation Deck: The presentation material used during the project is hosted online.
Presentation Video: A video summarizing the project is also available.
Codebase: The project’s code is hosted on Georgia Tech’s GitHub.
README: A README file is provided to guide the TAs through reproducing the results.

Link to presentation deck and video is provided here.

Abstract - This report explores the use of Motion History Images (MHI) and Hu Moments for classifying
human activities in video sequences. A complete pipeline is developed to extract binary motion
signals from grayscale video frames, compute MHIs, and derive primarily Hu moment-based features
for classification. The approach is evaluated on the KTH Human Action Dataset, focusing on
six activities: walking, jogging, running, boxing, hand-waving, and hand-clapping. Developed
system achieves competitive performance using a k-Nearest Neighbors (k-NN) classifier trained on
invariant moment features. Experimental results demonstrate both the strengths and limitations of
the approach. I also analyze classification errors due to inter-class similarities in motion patterns.

Abstract - This paper presents a series of experiments and analysis
for the hateful meme classification task. The dataset used in this
study is provided during Phase 1 of the Facebook Hateful Meme
Detection Challenge. Various uni-modal and multi-modal models,
including custom models, are included in this study. Feature
augmentation techniques are adopted for model performance
improvement. Observations from experiments are analyzed
iteratively and additional techniques are deployed accordingly to
pursue performance improvement.
Keywords - Machine Learning, Supervised, Compare Models,
Breast Cancer, Phishing

Abstract
Continuous and unobtrusive respiratory rate (RR) monitoring is
clinically valuable but is rarely deployed outside of intensive care,
where wearable or contact sensors can be uncomfortable and im-
practical for long-term use. We present a touchless RR monitoring
pipeline based on an infrared (IR) depth camera positioned above or
in front of the torso. The system automatically localizes the torso,
tracks respiration-driven surface motion under blankets, extracts a
breathing waveform, and computes instantaneous and 60-second
average RR in real time at 15 fps. Because the depth sensor relies
on active IR projection rather than ambient visible light, its per-
formance is largely invariant to room brightness and works in
complete darkness, unlike RGB camera-based methods. A per-pixel
respiratory-band energy map guides a probabilistic motion mask
that adapts to blankets and posture shifts; Hilbert-transform in-
stantaneous frequency estimation with Kalman smoothing (𝑄=0.05,
𝑅=2.0) yields stable, responsive RR estimates. An explicit weighted
confidence score and a dual-criterion apnea detector - amplitude
collapse and motion-blob area collapse, both sustained for 10 s with
3-second hysteresis - complete a clinician-oriented dashboard dis-
playing the depth view, light-green/amber breathing color overlay,
waveform, RR trends, and apnea flags. We demonstrate perfor-
mance across three progressively challenging conditions: without
a blanket, with a blanket, and with a blanket in total darkness.

In this work we took the approach to effectively track
alcohol and tobacco smoking consumption for self-management
by designing and developing a readily accessible web application
that provides individuals with the ability to log this information
quickly, track their vitals over time and see the trends of
consumption and vitals. We also ran population-level trend
analysis on generally available anonymized patient profile data,
observations and health conditions in a FHIR database to find
correlation between patient vitals, substance consumption and
chances of developing the health conditions, This insight is
integrated with the individual patient profile, vitals and lifestyle
choices they make for substance consumption to show the
increased chances of certain health conditions on substance
consumption. Individual inferences for increase in probable health
conditions from population data analysis can be improved with
more volume of patient data available in future.
Keywords - FHIR, Population level analysis, Alcohol and tobacco
smoking, Health conditions

Abstract— This paper covers observations and analysis during
and after running machine learning experiments on two different
public datasets. Observations by applying supervised machine
learners (widely known as Decision Tree, Adaptive Boosting,
Support Vector Machines, KNN and Artificial Neural Networks)
on the chosen dataset, are compared and contrasted.
Keywords— Machine Learning, Supervised, Compare Models,
Breast Cancer, Phishing

Abstract— This paper covers observations and analysis during
and after running experiments with model-based and model-free
techniques on two different Markov Decision problems.
Observations by applying these methods (Model Based - Value
iteration, policy iteration and Model Free - Q-learning) on the
chosen problems, are compared and contrasted. Methods are
tuned for the parameters to get the optimized results on best effort
and time bound basis. Results and behavior of these methods are
analyzed in different problem size environments.
Keywords— Model Based, Model Free, Reinforcement Learning,
Forest Management, Frozen Lake

Abstract— This paper discusses fundamentals of Temporal
Difference learning and compares and contrasts it with the
conventional learning at the time Sutton’88 was published. It
describes Random Walk experiments replicated from Sutton’88
[1] paper. It details the observations and analysis during and after
implementation of these experiments. Paper empirically augments
understanding of general conclusions made in the original paper.
Keywords— TD Learning, Sutton88, Random Walk, Temporal
Difference Prediction, Incremental Learning