Selected Research Projects
In my research career, I began with deep learning and optimization but eventually pivoted to reinforcement learning. The main reason for this shift was my passion for understanding the problem of intelligence, and reinforcement learning felt like the perfect field to explore this exciting challenge.
We investigate the properties of representations learned by deep reinforcement learning systems for transfer learning, measuring six key properties across 25,000 agent-task settings with Deep Q-learning and Rainbow agents to understand their correlation with transfer performance.
We investigate reinforcement-learning-based prediction for a drinking-water treatment plant, addressing challenges like seasonality and partial observability. We show that General Value Function (GVF) predictions and temporal difference learning agents outperform n-step predictions, highlighting the importance of real-time adaptation in non-stationary systems.
We propose a novel transfer evolutionary optimization framework with two co-evolving species to efficiently handle over 1000 source tasks, demonstrating scalability and effective knowledge capture despite source sparsity in large-scale cloud-based applications.