By Ibrahim Alwishah
Would you like to take part in AI research exploring the different machine learning algorithms and methodologies?
My response to this question was an obvious yes. It is not common for an undergraduate student to take part in such advanced research involving the developing world of AI. My introduction to DAgger came when I introduced myself to Professor Das before being chosen to take part in the summer research project. During that same semester, I took an Intro to AI class and was very intrigued by the topics we had studied during that class. When Professor Das approached me about the intriguing potential of the decision-making systems I would be working with, I immediately accepted the opportunity. I knew this opportunity would be an invaluable addition to my academic journey.
Our research aims to push the boundaries of reinforcement learning by leveraging methodologies like DAgger and HG-DAgger as well as variants of these methodologies in the future. We specifically explored how these algorithms could enhance AI decision-making by integrating pseudo-human expert feedback. As a computer science student, I knew that gaining experience in this field would deepen my knowledge and contribute to my career by building a knowledge base to work in this field in my future career.
We built our experiment in an OpenAI Gym environment, training a pseudo-human agent using Deep-Q-Learning (DQN) to act as an expert agent. The agent works on iteratively guiding a less experienced AI, allowing us to refine its decision-making through the DAgger framework. The process was very interesting watching the AI’s performance improve after every iteration and tracking the increase in rewards while noting improvement with the expert agent’s guidance was a fulfilling experience.
My involvement in this project was part of the Summer Undergraduate Research Experience (SURE) Program, this program offered me the opportunity to begin research at the university which I had never thought of before. The program offered guidance and support throughout the research process, overall refining my research skills and allowing me to continue research at the university. The highlight of the program was getting to present our preliminary findings at the SURE Showcase event.
Looking forward, we aim to test different variations of DAgger, exploring uncertainty estimates and possibly incorporating newer AI models. The research is still in its early stages but we are already envisioning the potential it has in shaping the future of decision making in different autonomous systems.
