franktiantt / intact

Identifiable aNd TrAnsferable Causal meTa World Model.

INTACT

INTACT stands for Identifiable aNd TrAnsferable Causal meTa world model, aiming to address the context decoupling issue in meta reinforcement learning. Its goal is to restore fully recognizable context representations using observed data while ensuring swift adaptation to new contexts within the target environment distribution without sacrificing the original semantics.

Installation

Using python 3.11 for example:

# create conda env
conda create -n intact python=3.11
conda activate intact

# install torch, torchvision, tensordict
pip3 install torch torchvision --index-url https://download.pytorch.org/whl/cu118
pip install tensordict-nightly -i https://pypi.org/simple # do not use mirror

# install intact
https://github.com/FrankTianTT/intact.git --depth 1
cd intact
pip install -e .

Using

Dreamer in POMDP:

xvfb-run --auto-servernum --server-num=1 python examples/dreamer/train.py

Dreamer in MDP:

python examples/dreamer_mdp/train.py

MPC in MDP:

python examples/mdp/train.py

Design