This is an implementation of distributed deep reinforcement learning framework called Importance-Weighted Actor-Learner Architectures (Impala) using DeepMind Lab simulator and distributed execution framework Ray. Currently, only 1-step importance sampling is implemented, V-trace will be implemented later.
To run the program (actor/learner/parameter server) use the following command: python learner.py --length 100 --actors 10 -s --level_script
Set IP address of RAY HEAD in learner.py before running, if running Ray configured as a cluster.