PyTorch Implementation of Ape-X (Distributed prioritized experience replay) architecture with DQN learner
- Easy-to-follow implementation with comments indicating the algorithm line as described in the paper
-
Setup a conda env with the necessary python packages. Assuming Anaconda is installed, you can run the following command from the root of this repository:
conda env create -f conda_env.yaml -n "apex_dqn_pytorch"
-
Set the configuration parameters suitable for your hardware in parameters.json At the minimum, you should set the
num_actors
parameter under"Actor"
and the"Replay Memory"
'soft_capacity` based on the number of CPU cores and RAM available. -
Launch the training process using the following command:
python main.py
You should see episode stats printed out to the console. You can change the learning environment
using the "name"
parameter value under "env_conf"
in parameters.json
- Compress state/observations using PNG codec before storing in memory and decompress when needed
- Bias correction in prioritized replay using importance sampling