All Posts
Category
Reinforcement Learning
2 posts
-
Teaching a Snake AI to See Its Own Tail: Why Representation Beats Raw Compute
How adding a flood-fill observation to a DQN Snake agent made 100k training steps outperform 2.5 million — because the agent was blind, not dumb.
-
Three Iterations on a 2048 AI: What Representation and Reward Design Actually Do
How upgrading a 2048 DQN agent from raw score rewards and a flat MLP to one-hot encoding, a CNN, and a multi-component reward function changed what the agent could actually learn.