yiqinyang/iac

Codes accompanying the paper "Modeling the Interaction between Agents in Cooperative Multi-Agent Reinforcement Learning" (AAMAS 2021 https://arxiv.org/abs/2102.06042)

IAC

This is a pytorch implementation of IAC on Multi-Agent Particle Environment(MPE), the corresponding paper of IAC is Modeling the Interaction between Agents in Cooperative Multi-Agent Reinforcement Learning.

Requirements

• python=3.6.5
• Multi-Agent Particle Environment(MPE)
• torch=1.1.0

Quick Start

$ python3 main.py --scenario-name=simple_tag --evaluate-episodes=10

Directly run the main.py, then the algrithm will be trained on scenario 'simple_tag' for 10 episodes.

Note

• The POMDP version of simple_tag is in POMDP_tag.py. You can put it in the MPE environment and enjoy it.

• There are 4 agents in simple_tag, including 3 predators and 1 prey. we use IAC to train predators to catch the prey. The prey's action can be controlled by you, in our case we set it random.

• The default setting of Multi-Agent Particle Environment(MPE) is sparse reward, you can change it to dense reward by replacing 'shape=False' to 'shape=True' in file multiagent-particle-envs/multiagent/scenarios/simple_tag.py/.

• Our work is basic, and I think someone can explore some exciting directions based on this work. If you have any questions, please contact me: [email protected].

iac's Issues

Regarding eval_mode

IAC/IAC/actor.py

Line 39 in 925dc26

def forward(self, x, train=False, eval_mode=False):

Hi, Aravind here. During decentralized execution, it looks like eval_mode is still set to False. Does this mean that you are using the entire global state for action selection during execution also?