Skip to content

Latest commit

 

History

History
21 lines (14 loc) · 947 Bytes

README.md

File metadata and controls

21 lines (14 loc) · 947 Bytes
Raw

Distributed reinforcement learning with pytorch RPC

Here is an example on how to run distributed_per_sac.py on two computers. The learner and one actor is run on node with IP 192.168.178.101, network interface eth0:

export GLOO_SOCKET_IFNAME=eth0 && export TP_SOCKET_IFNAME=eth0 && python distributed_per_sac.py --world-size 3 --learner-addr 192.168.178.101 --learner-port 8080 --rank 0

and

export GLOO_SOCKET_IFNAME=eth0 && export TP_SOCKET_IFNAME=eth0 && python distributed_per_sac.py --world-size 3 --learner-addr 192.168.178.101 --learner-port 8080 --rank 1

Another actor is run on another node, with network interface eth1:

export GLOO_SOCKET_IFNAME=eth1 && export TP_SOCKET_IFNAME=eth1 && python distributed_per_sac.py --world-size 3 --learner-addr 192.168.178.101 --learner-port 8080 --rank 2

This script will use soft actor-critic algorithm with distributed prioritized experience replay to learn.