I grew interested in reinforcement learning after reading DeepMind's paper on how they made a single agent that could learn to competently play 6 different Atari games using raw video input. Their approach was simple but brilliant: they took the standard Q-learning algorithm and swapped its lookup table with a convolutional neural network. This allowed Q-learning to be applied to an infinite and continous state space.
To cement my understanding of the algorithm, I applied it to the classic Snake game. I used Keras for the convolutional neural net and trained it in a custom OpenAI Gym environment. And of course, I used Weights & Biases to track my experiments.
I wasn't able to perfectly solve Snake like I had hoped, but it's still pretty neat to watch. I'm now running the trained model in the background of my home page. It's secretly interactive; you can temporarily override the AI's choices using the arrow keys.