CS470RL/maac-pettingzoo

Applying the MAAC algorithm onto Pettinzoo Environments

<<<<<<< HEAD

Multi-Actor-Attention-Critic

Save average rewards for each agent per episode, and plot 'avg rewards per episode'

• Paper Actor-Attention-Critic for Multi-Agent Reinforcement Learning (Iqbal and Sha, ICML 2019)
• Original code repo : https://github.com/shariqiqbal2810/MAAC

How to run MAAC code & plot

python main.py fullobs_collect_treasure mytest1 --n_episodes 10000 --n_rollout_threads 1 --testnum 1

python plot.py --input test1.csv --which 0

• main.py --testnum 'testnum' option in main.py MAAC code saves the rewards to test{testnum}.csv
  plot.py takes 'test{testnum}' or 'test{testnum}.csv' as an input using --input option.

• main.py --n_rollout_threads 1
  🚨 currently only supports single process

• plot.py --which agent to plot
  plot.py takes the agent number to plot with --which option.
  0 to plot every agents, or 1~NUMOFAGENTS to plot individual agent (NUMOFAGENTS in MAAC paper is 8)

Experiment Results

• env: fullobs_collect_treasure
• n_episodes: 10,000
• n_rollout_threads: 10

Training statistics for a single agnet (agent0)

Total sum of rewards in each episode

maac-pettingzoo

Applying the MAAC algorithm onto Pettingzoo Environments

06f27286cc120afd90dcf43d223ab0a9b844643f