Diana Huang, Shalini Keshavamurthy, and Nitin Viswanathan
This repository contains the code used in our CS234 Winter 2018 class project, where we implemented both MADDPG and policy gradient approaches to train agents in a cooperative navigation scenario.
-
pg/config.py, pg/pg.py: contains code for the configuration and implementation of running multiple independently learning agents using a single-agent policy gradient approach -
maddpg/config.py, maddpg/maddpg.py, maddpg/ddpg_actor_critic.py: contains code for the configuration and implementation of running multiple agents learning using MADDPG -
.main.py: Modify this to determine which algorithm you run (vanilla PG vs. MADDPG) -
./utils/: various utility functions we wrote to run the multi-agent scenarios
-
Known dependencies: tensorflow, OpenAI gym, numpy, OpenAI multi-agent-envs(https://github.com/openai/multiagent-particle-envs)
-
ensure OpenAI multi-agent-envs is cloned to the root directory of this project
-
simply run
python main.pyto begin running with our default settings. -
if you want to change the algorithm between PG and MADDPG, edit main.py
-
if you want to change experiment settings, edit the appropriate config.py file