Aarshdeep Singh N's Senior Project Blog
 |
Project Title: Interpreting Reinforcement Learning Agents BASIS Advisor: Camille Bennett Internship Location: University of Arizona Onsite Mentor: Dr. Chicheng Zhang, Assistant Professor, Computer Science |
Project Abstract
This project aims to develop an interpretable Reinforcement Learning (RL) framework that enables AI agents to explain their decision-making processes in a way that is understandable to humans. While RL has achieved remarkable success in tasks ranging from game playing to robotics, the ""black-box"" nature of these systems often makes it difficult to understand why an agent chooses a specific action. This lack of transparency limits trust, hinders debugging, and restricts the deployment of RL in critical applications where explainability is essential. The goal of this research is to create an AI system that not only performs tasks effectively but also provides clear, actionable explanations for its actions, helping users understand what the AI is ""thinking"" when it makes a decision. The project will involve training RL agents using state-of-the-art algorithms (e.g., Deep Q-Networks, Proximal Policy Optimization) in diverse environments, such as grid worlds, Atari games, or simulated robotics tasks. To interpret the agent's behavior, we will integrate explainability techniques such as saliency maps (to highlight important features in the input), counterfactual analysis (to explore alternative actions), and natural language generation (to translate the agent's reasoning into human-readable explanations). For example, the system might explain, ""The agent moved left to avoid an obstacle and maximize long-term rewards."" The final product will be a modular framework or tool that can analyze and interpret the actions of RL agents, providing insights into their decision-making processes. This research has broad applications, from explaining AI behavior in games like Chess or Go to interpreting the decisions of autonomous systems like self-driving cars or drones. By making RL agents more transparent and interpretable, this project will enhance trust, facilitate debugging, and enable safer deployment of AI in real-world scenarios. The end goal is to bridge the gap between AI performance and human understanding, making RL more accessible and reliable for both researchers and practitioners.
Week 8 updates
This week has been all about taking our project to the next level. First, we upgraded our decision boundary visualization—moving from a 2D graph to a 3D decision boundary graph inspired by a method from a StackOverflow post. This allowed us to plot and understand three state variables at once, which has already helped us... Read More
Week 7 Update
This week has been a mix of reflection, correction, and continued learning. After reviewing my progress from last week, I realized that I had made a mistake with the parameters in the decision boundary graph I created. I had accidentally used values of +24 and -24 for the pole angle range, which led to incorrect... Read More
Week 6
Hello everyone, and welcome back to Week 6 of my Senior Research Project on Reinforcement Learning! This week was all about brainstorming and defining the direction for the end product of our project. We focused on how to present our work in the most meaningful and impactful way, especially considering the complex nature of reinforcement... Read More
Week 5!
Hello everyone, and welcome back to Week 5 of my Senior Research Project on Reinforcement Learning! This week, I spent more time exploring complex variations of the CartPole problem, pushing beyond the basic reinforcement learning models I used before. I started researching Deep Q-Learning (DQL) and how neural networks play a role in improving decision-making.... Read More
Week 4 Updates
Hello everyone, and welcome back to Week 4 of my Senior Research Project on Reinforcement Learning! This week was an exciting step forward—I finally ran my own reinforcement learning simulation using OpenAI’s Gym environment and trained an agent to balance a pole in the classic CartPole problem. I spent time going through multiple YouTube tutorials... Read More
Week 3 Updates!
Hello everyone, and welcome to Week 3 of my Senior Research Project on Reinforcement Learning! This week was a bit different due to the senior trip, which meant I didn’t have as much time for hands-on coding. Instead, I focused on deepening my theoretical understanding of reinforcement learning by reading through a fascinating research paper... Read More
Week 2
Hello everyone, and welcome to my week 2 updates of the senior research project. After spending my first week learning the fundamentals of reinforcement learning, this week I shifted my focus toward finding open-courseware and resources to start training my own RL model. The transition from theory to implementation has been exciting, as I explored... Read More
Blog Post Week 1
Hi everyone, welcome to my Reinforcement Learning Research Project! During Week 1, I took the Fundamentals of Reinforcement Learning course on Coursera to build a strong foundation in RL concepts. I began by understanding the key differences between supervised learning and reinforcement learning. Unlike supervised learning, where a model is trained with labeled data, reinforcement... Read More
Introduction
Hello Everybody, My name is Aarsh, and I am currently working on a Senior Research Project in Explainable AIs. My interest in AI, particularly in autonomous systems, stems from a fascination with how intelligent systems can solve complex real-world problems. While I initially aimed to pursue autonomous systems research in college, circumstances prevented me from... Read More