GCM-DQN

Setup

conda env create -f venv.yml
conda activate gcm-dqn
python -m unittest discover -s tests -v

Expected dataset layouts are:

data/geom_data/<domain>/plans/*
data/geom_data_test/<domain>/plans/*
data/ipc_short/<domain>/Jax/plans/<instance>/*
data/ipc_short_test/<domain>/Jax/plans/<instance>/*

Run experiments

The main paper configuration uses a separate stochastic transition model for each action. For example:

python code/main.py --task experiment --planner gcmdqn \
  --stm stoch_single --domain circle --seed 1 --device cuda

Replace gcmdqn with pdqn or pddpg for a baseline. IPC runs additionally need an instance:

python code/main.py --task experiment --planner gcmdqn \
  --stm stoch_single --domain HVAC_ippc2023 --instance instance0 \
  --seed 1 --device cuda