Code for a MineRL agent. Runs on Arch Linux, Python 3.8.5.
pip install -r requirements.txt
Get the dataset (60GB) and install in data/: https://minerl.io/dataset/
To run a training loop, then evaluate:
python main.py --train
11 discrete actions. Total list of actions:
- cam left
- cam right
- cam up
- cam down
- place + jump
- place
- forward + attack
- attack
- forward + jump
- jump
- forward
For Curriculum Learning, we learn a few actions at a time. We do this by keeping the same action space, but masking actions we don't want to learn yet by multiplying by zero.
- cam left
- cam right
- forward
- cam left
- cam right
- forward + jump
- jump
- forward
- cam left
- cam right
- cam up
- cam down
- attack
- forward + jump
- jump
- forward
- cam left
- cam right
- cam up
- cam down
- place + jump
- place
- forward + attack
- attack
- forward + jump
- jump
- forward