Atari, pokemon, Monoply and Dark Souls (or, I really want to get back to RL for games)

I think the coolest type of ML personal projects are the type where we are doing game strategy discovery through AI, or RL to train bots. I think if resources allow, there’s no better bug checker than a RL system. Atari gyms (Atari 100k and its ilk) are well known “benchmarks”, but in this particular case, the codebase is great and the training efficiency is great. It’s also a semi survey paper to ge into the history of techniques used for this benchmark.

The Pokemon and Monoply examples are super fresh. They represent the two directions I have stated: in the pokemon one we are building a human level agent and in the monoply one, we are (re) discovering strategies. I think this is the absolutely best part of these game RL agents: both create insights and create interaction, much more so than the standard supervised and unsupervised projects.

The dark souls one seems a bit dead, but let me see whether I can find some code for it! Worse come to worse we always have this tutorial to work on. Sentdex is an absolute legend.

Best place to start: stable retro. It’s Gym, but with retro games