This library unifies the distribution correction estimation algorithms for off-policy evaluation, including:
- DualDICE: Behavior-Agnostic Estimation of Discounted Stationary Distribution Corrections
- GenDICE: Generalized Offline Estimation of Stationary Values
- Reinforcement Learning via Fenchel-Rockafellar Duality Please cite these work accordingly upon using this library.
Existing DICE algorithms are the results of particular regularization choices in the Lagrangian of the Q-LP and d-LP policy values.
Choices of regularization (colored) in the Lagrangian.
These choices navigate the trade-offs between optimization stability and estimation bias.
Estimation bias given the choices of regularization.
Navigate to the root of project, and perform:
pip3 install -e .
To run taxi, download the pretrained policies and place them under policies/taxi:
git clone https://github.com/zt95/infinite-horizon-off-policy-estimation.git
cp -r infinite-horizon-off-policy-estimation/taxi/taxi-policy policies/taxi
First, create datasets using the policy trained above:
for alpha in {0.0,1.0}; do python3 scripts/create_dataset.py --save_dir=./tests/testdata --load_dir=./tests/testdata/CartPole-v0 --env_name=cartpole --num_trajectory=400 --max_trajectory_length=250 --alpha=$alpha --tabular_obs=0; done
Run DICE estimator:
python3 scripts/run_neural_dice.py --save_dir=./tests/testdata --load_dir=./tests/testdata --env_name=cartpole --num_trajectory=400 --max_trajectory_length=250 --alpha=0.0 --tabular_obs=0
To recover DualDICE, append the following to the above python command:
--primal_regularizer=0. --dual_regularizer=1. --zero_reward=1 --norm_regularizer=0. --zeta_pos=0
To recover GenDICE, append the following to the above python command:
--primal_regularizer=1. --dual_regularizer=0. --zero_reward=1 --norm_regularizer=1. --zeta_pos=1
The configuration below generally works the best:
--primal_regularizer=0. --dual_regularizer=1. --zero_reward=0 --norm_regularizer=1. --zeta_pos=1