Thatโs a wrap on our GSoC session with Diya Srivastava, GSoCโ26 Contributor.
An insightful conversation, a curious crowd, and plenty to take away.
More coming from ELIXIR.
GSoC sounds exciting. Getting there is a different story.
This Saturday, join ELIXIR for a conversation with Diya Srivastava, GSoCโ26 JSON Schema, as she shares the journey, lessons, and what it really takes to get there.
Saturday | 8 PM
Discord: https://t.co/kgbUBVvMEx
Training a Pong RL agent from scratch. ๐
Started with a custom physics-based Pong environment and a deterministic geometric AI that predicts the ballโs future impact position using its velocity, trajectory, and wall reflections. It was almost UNBETABLE
state/action space, epsilon-greedy exploration, experience replay,target network, Bellman updates, and evaluation pipeline from scratch.
Current milestone: 65%win rate after 100K training steps against a controlled opponent.