Multiagent Reinforcement Learning (MARL) is increasingly critical for real-world applications, yet benchmarking environments that capture the complexities of these scenarios remain a challenge. The Multi-Agent Tracking Environment (MATE) introduces a novel solution by simulating target coverage control, featuring an asymmetric cooperative-competitive game. This environment pits two distinct agent groups against each other: “cameras” and “targets”. Cameras, acting as directional sensors, are tasked with maximizing the coverage of targets, while targets are mobile agents aiming to transport cargo and minimize their exposure to the camera network.
MATE is designed to rigorously benchmark MARL algorithms across several key dimensions: cooperation, communication, scalability, robustness, and asymmetric self-play. Initial benchmarks within MATE have employed established MARL algorithms such as MAPPO, IPPO, QMIX, and MADDPG for cooperative tasks. Furthermore, the integration of multi-agent communication protocols like TarMAC and I2C has been explored to enhance agent coordination. To evaluate competitive dynamics, self-play techniques including PSRO and fictitious self-play have been applied, fostering a co-evolutionary process between cameras and targets. Notably, the emergence of diverse roles among target agents has been observed when incorporating I2C for inter-target communication.
Implemented purely in Python and compatible with the OpenAI Gym API, MATE offers a user-friendly platform for researchers. This benchmark facilitates deeper investigation into Asymmetric Multiagent Reinforcement Learning, providing valuable insights into developing robust and adaptable multi-agent systems. The project is publicly available, encouraging further research and development in this vital area of AI.
