Celeste-AI: A Celeste Classic DQL Agent

This is an attempt to create a deep Q-learning agent that automatically solves the first stage of Celeste Classic.

A gif of the result is below. This took 4000 episodes, which amounts to about 30 hours of training time.

Contents

• ./resources: contains files this script requires. Notably, we have an (old) version of PICO-8 that's known to work with this script, and a version of Celeste Classic with telementery and delays called hackcel.p8.
• ffmpeg.sh: uses game screenshots to make real-time video of the agent's attempts. Read the script, it's pretty simple.
• plot.py: generates plots from model snapshots. These are placed in model_data/current/plots/.

Setup

Before you set up Celeste-AI, you need to prepare PICO-8. See resources/README.md

This is designed to work on Linux. You will need xdotool to send keypresses to the game.

1. cd into this directory
2. Make and enter a venv
3. pip install -e .

Once you're set up, you can...

• python celeste_ai/train.py to train a model
• python plot.py to make prediction plots
• python test.py to test a model

Before running, be aware of the following:

• Only one instance of PICO-8 can be running at a time. See celeste.py.
• hackcel.p8 captures a screenshot of every frame. PICO-8 will probably place these on your desktop. Since this repo contains a rather old version of PICO-8, there is no way to change where it places screenshots. train.py will delete, move, and rename screenshots automatically during training, but you should tell it where your desktop is first.
• When you start training, a model_data directory will be created. It contains the following:
  • model_archive: history of the model. Save interval is configured inside train.py
  • screenshots: contains subdirectories. Each subdirectory contains the frames of one episode. Use ffmpeg.sh to turn these into a video.
  • plots: generated by plot.py. Contains pretty plots.