Deep RL agent for Pong
Aug 15, 2024
·
1 min read
During my internship at DISCOVERY LAB GLOBAL, I trained an autonomous agent to play and win Pong using deep reinforcement learning and convolutional neural networks in Python.
What I built
- A custom simulation to log rewards, frame stacks, and policy rollouts so the team could compare runs quickly.
- Training loops with tuned discount factors, learning rates, and exploration schedules to stabilize policy improvement.
- A concise technical paper capturing architecture choices, experiments, and lessons learned for stakeholders.
Outcomes
The project sharpened my intuition for credit assignment, sample efficiency, and how engineering choices in the simulator change what the policy can learn—skills I still apply when thinking about data quality and evaluation for ML systems.
Authors
CSE & Math @ OSU · Robot Operator @ Diligent Robotics
Computer Science and Engineering student at The Ohio State University focused on machine learning and systems programming in Python and C++. I care about shipping reliable software—from deep reinforcement learning experiments to real-world robotics in hospitals—and about learning from every layer of the stack.