- This is a pytorch implementation developed for Return-Gap-Minimization Communication(RGMComm) algorithm
- It contains two folder: RGMComm_stage1 and RGMComm_stage2_stage3, see more detailed readme files under each folder:
- STAGE 1 to collecting action-value(Q values) vectors samples from trained centralized critic;
- STAGE 2 and STAGE 3 to train message generation module and train agents with communication message labels enabled;
- Evaluation environment is Multi-Agent Particle Environment(MPE), the corresponding paper is Multi-Agent Actor-Critic for Mixed Cooperative-Competitive Environments.
- python=3.6.5
- Multi-Agent Particle Environment(MPE)
- torch=1.1.0