This repository houses code for the AAAI 2021 paper:

GLIB: Efficient Exploration for Relational Model-Based Reinforcement Learning via Goal-Literal Babbling

Rohan Chitnis*, Tom Silver*, Joshua Tenenbaum, Leslie Pack Kaelbling, Tomás Lozano-Pérez.

For any questions or issues with the code, please email ronuchit@mit.edu and tslvr@mit.edu.

Link to paper: https://arxiv.org/abs/2001.08299

Instructions for running:

• Use Python 3.5 or higher, e.g. with a virtual environment.
• Download Python dependencies: pip install -r requirements.txt.
• Download the Fast-Forward (FF) planner to any location on your computer. -> Linux: https://fai.cs.uni-saarland.de/hoffmann/ff/FF-v2.3.tgz -> Mac: https://github.com/ronuchit/FF
• From the FF directory you just created, run make to build FF, producing the executable ff.
• Create an environment variable "FF_PATH" pointing to this ff executable.
• Back in the GLIB directory, you can now run python main.py.

By default, the code runs GLIB-L, GLIB-G, Oracle, and Action babbling (also called "random") on the Blocks domain. If you want another domain or only some of the methods, change domain_name or curiosity_methods_to_run in settings.py. Plots will get written out after each seed into an automatically created results/ folder. Here is an example of the rough shape of plots that should result from running this code out-of-the-box (it may take around 15 minutes to complete):