phillip

Why we included this project

Phillip is a research-grade deep reinforcement learning agent that learns to play Super Smash Bros. Melee by interacting with the Dolphin emulator, and it is one of the more complete public examples of training a competitive fighting-game bot end to end. Anyone curious about how game agents are built from raw emulator state rather than hand-coded rules will find a full pipeline here: a Python trainer, parallel training launchers for Slurm clusters, TensorBoard logging, and a set of pretrained agents you can load and watch play. The project is no longer maintained and expects a specific NTSC 1.02 ISO plus a custom Dolphin build, so treat it as a reference implementation and a starting point rather than a drop-in tool. For teams exploring reinforcement learning in games, the value is in the concrete, readable example of the architecture and training loop; the README also points to a successor project using imitation learning from Slippi replays.