This repository houses a minimal PyTorch implementation of Implicit Q-Learning (IQL), an offline reinforcement learning algorithm, along with a script to run IQL on tasks from the D4RL benchmark.

To install the dependencies, use pip install -r requirements.txt.

You can run the script like so:

python main.py --log-dir /path/where/results/will/go --env-name hopper-medium-v2 --tau 0.7 --beta 3.0

Note that the paper's authors have published their official implementation, which is based on JAX. My implementation is intended to be an alternative for PyTorch users, and my general recommendation is to use the authors' code unless you specifically want/need PyTorch for some reason.

I am validating my implementation against the results stated in the paper as compute permits. Below are results for the MuJoCo locomotion tasks, normalized return at the end of training, averaged (+/- standard deviation) over 3 seeds:

We can see that the performance is mostly similar to what is stated in the paper, but slightly worse on a few tasks. Note that these results were obtained using a small simplification (deterministic policy and least-squares loss rather than a Gaussian distribution and negative log likelihood), which may explain the discrepancy.