README.md

Skipper

A PyTorch Implementation of Skipper, proposed in the ICLR 2024 paper

Consciousness-Inspired Spatio-Temporal Abstractions for Better Generalization in Reinforcement Learning

-- Mingde Zhao, Safa Alver, Harm van Seijen, Romain Laroche, Doina Precup, Yoshua Bengio

arXiv

blogpost

Python virtual environment configuration:

1. Create a virtual environment with conda or venv (we used Python 3.9)

2. Install PyTorch according to the official guidelines, make sure it recognizes your accelerators

3. pip install -r requirements.txt

For experiments, write bash scripts to call those Python files that start with string "run_":

run_minigrid_mp.py: a multi-processed experiment initializer for Skipper agents.

run_minigrid.py: a single-processed experiment initializer for modelfree baseline

run_minigrid_with_CVAE.py: a single-processed experiment initializer for training a checkpoint generator with the experience colleced by a modelfree or random baseline

run_leap_pretrain_vae.py: a single-processed experiment initializer for pretraining generator for the adapted LEAP agent

run_leap_pretrain_rl.py: a single-processed experiment initializer for pretraining distance estimator (policy) for the adapted LEAP agent

Please read carefully the args definition in runtime.py and pass the desired args in the commands.

Extras

• There is a potential CUDA_INDEX_ASSERTION error that could cause hanging at the beginning of the Skipper runs. We don't know yet how to fix it
• The Dynamic Programming solutions for environment ground truth are only compatible with deterministic experiments