Skip to content

Code for "Principled Offline RL in the Presence of Rich Exogenous Information" and "Ignorance is Bliss: Robust Control via Information Gating".

Notifications You must be signed in to change notification settings

manantomar/agent-centric-representations

Repository files navigation

Usage

The code contains implementations of two algorithms

  1. ACRO - learns agent centric representations with a multi-step inverse dynamics model.
  2. InfoGating - adds an InfoGating bottleneck over a contrastive variant of the multi-step inverse dynamics model.

Installation

The provided requirements.txt file contains all main dependencies required to run this code (tested on Python 3.8). You will also need to download the offline datasets from v-d4rl paper or (depending on which kinds of distractors you want to test with) from the ACRO paper and have them in ./vd4rl path.

Running

Simply use the train.py script to run ACRO/InfoGating. Provide a task_name, offline_dir directory where the offline code is available and which algo to run. The dist_level argument is only used for naming result files.

Run ACRO

python train.py task_name=offline_cheetah_run_expert offline_dir=/path/to/dataset/vd4rl/main/cheetah_run/expert/ seed=1 algo=acro dist_level=none

Run InfoGating

python train.py task_name=offline_cheetah_run_expert offline_dir=/path/to/dataset/vd4rl/main/cheetah_run/expert/ seed=1 algo=infogating dist_level=none

About

Code for "Principled Offline RL in the Presence of Rich Exogenous Information" and "Ignorance is Bliss: Robust Control via Information Gating".

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages