My research sits at the crossroads of artificial intelligence, human-computer interaction, and economic systems. I am particularly drawn to questions about how intelligent agents, human and artificial, coordinate under uncertainty.
Multi-Agent Reinforcement Learning in Ridesharing Systems
In this project, I studied how autonomous agents can learn cooperative and competitive strategies in a simulated ridesharing environment. Using standard RL algorithms, I analyzed reward dynamics, convergence behavior, and the economic implications of different incentive structures. The work resulted in a Shiny dashboard for visual exploration of simulation results and configuration comparisons.
Key questions: How do individual agent strategies affect system-wide efficiency? Can shared-reward structures outperform individual-reward models in multi-agent logistics?
Predicting Opioid Overdose Death Rates Across U.S. Counties Using Supervised Machine Learning
The opioid crisis in the United States has been one of the most significant public health emergencies of the past several decades. Fentanyl, a synthetic opioid 50 times more potent than heroin, contributed to around 31,000 deaths, comprising 65% of all opioid overdose fatalities by 2018. This research trains and evaluates multiple supervised machine learning models using unrestricted datasets spanning 2012 to 2021 to identify the most effective predictive model for opioid overdose death rates across U.S. counties.
The ultimate goal is to use these predictions as a guide for strategic decision-making in the investment and allocation of public health services, particularly in hotspot counties, to proactively prevent future risk. The study also investigates the social, economic, and demographic features influencing these outcomes, aiming to prioritize and understand their significance.
Dataset & scope: Unrestricted county-level data across the contiguous United States (excluding Alaska and Hawaii), 2012–2021. Predictors include health-care access, drug market indicators, socio-demographics, and geographical distribution of opioid overdoses.
Methodology:
- Trained six different supervised ML models on data from 2012 to five years before the predicted year.
- Converted death count outputs into overdose deaths per 100,000 people for standardized comparison.
- Compared model predictions for the target year against actual observed rates using five evaluation metrics.
Key finding: Among the six trained models, the random forest exhibited the best performance given the available data, enhancing our understanding of the complex dynamics driving opioid overdose outcomes.
