This is a source code for Neural Experience Replay Sampler (NERS) which accepted in International Conference in Learning Representation 2021 (https://openreview.net/forum?id=gJYlaqL8i8).
Although NERS can be used in various algorithms, this repository applys Rainbow implemented in https://github.com/Kaixhin/Rainbow/blob/master/README.md
Run the following command to test NERS:
sh run.sh
The shell file will test NERS under data-effieicnt Rainbow (Prioritized Experience Replay (PER) is also used by "--replay-type PER"):
CUDA_VISIBLE_DEVICES=0 python main.py --target-update 2000 \
--T-max 100000 \
--learn-start 1600 \
--memory-capacity 100000 \
--replay-frequency 1 \
--multi-step 20 \
--architecture data-efficient \
--hidden-size 256 \
--learning-rate 0.0001 \
--evaluation-interval 1000 \
--replay-type NERS \
--game ms_pacman \
--enable-cudnn
To install all dependencies, run the following command:
pip install -r requirements.txt
