This is the repository for paper:

Towards Unbiased Action Value Estimation in Reinforcement Learning

The training requires WANDB to log data

To train a agent, go to ./shell_scripts and run:

./train.sh algo_name

algo_name is to be replaced by the following methods:

dvqn, dqn, sarsa, ddqn, cddqn, avgdqn, dueldqn

Hyperparameters are to be modified within train.sh

• trainig over Toy MDP, see ./toy_mdp/compare_TD_algos.ipynb
• training over 2-room gridworld, see ./gridworld/two_room.ipynb