Official implementation of ACC, described in the paper "Adaptively Calibrated Critic Estimates for Deep Reinforcement Learning". The source code is based on the pytorch implementation of TQC, which again is based on TD3. We thank the authors for making their source code publicly available.

1. Download and install MuJoCo 1.50 from the MuJoCo website. We assume that the MuJoCo files are extracted to the default location (~/.mujoco/mjpro150).

2. Copy your MuJoCo license key (mjkey.txt) to ~/.mujoco/mjkey.txt:

We recommend to use an anaconda environment. In our experiments we used python 3.7 and the following dependencies

pip install gym==0.17.2 mujoco-py==1.50.1.68 numpy==1.19.1 torch==1.6.0 torchvision==0.7.0

You can run ACC for TQC on one of the gym continuous control environments by calling

python main.py --env "HalfCheetah-v3" --max_timesteps 5000000 --seed 0

To run the data efficient variant with 4 critic update steps per environment step you can call

python main.py --env "HalfCheetah-v3" --max_timesteps 1000000 --num_critic_updates 4 --seed 0

An example script that runs the experiments for 10 seeds and all environments is in run_experiment.sh and run_experiment_data_efficient.sh.

You can speed up the experiments by using fewer networks in the ensemble of TQC. This trades off a little bit of performance for a faster runtime (see the Appendix of the paper). The number of networks can be controlled with the flag --n_nets. For example

python main.py --env "HalfCheetah-v3" --max_timesteps 5000000 --n_nets 2--seed 0