9 Dec 2019 | Mikayel Samvelyan, Tabish Rashid, Christian Schroeder de Witt, Gregory Farquhar, Nantas Nardelli, Tim G. J. Rudner, Chia-Man Hung, Philip H. S. Torr, Jakob Foerster, Shimon Whiteson
The paper introduces the StarCraft Multi-Agent Challenge (SMAC) as a benchmark for cooperative multi-agent reinforcement learning (MARL). SMAC is based on the popular real-time strategy game StarCraft II, focusing on decentralized micromanagement tasks where each unit is controlled by an independent agent. The challenge aims to evaluate how well agents can learn to coordinate their behavior under partial observability and high-dimensional inputs. SMAC offers a diverse set of scenarios that require agents to handle complex combat situations, such as focusing fire, kiting, and terrain exploitation. The paper also provides recommendations for benchmarking and evaluation practices, including standardized performance metrics and computational requirements. Additionally, it opens-source PyMARL, a deep MARL framework that includes state-of-the-art algorithms like QMIX, COMA, and IQL. The results demonstrate the performance of these algorithms on various SMAC scenarios, highlighting the need for more advanced coordination and exploration techniques in MARL.The paper introduces the StarCraft Multi-Agent Challenge (SMAC) as a benchmark for cooperative multi-agent reinforcement learning (MARL). SMAC is based on the popular real-time strategy game StarCraft II, focusing on decentralized micromanagement tasks where each unit is controlled by an independent agent. The challenge aims to evaluate how well agents can learn to coordinate their behavior under partial observability and high-dimensional inputs. SMAC offers a diverse set of scenarios that require agents to handle complex combat situations, such as focusing fire, kiting, and terrain exploitation. The paper also provides recommendations for benchmarking and evaluation practices, including standardized performance metrics and computational requirements. Additionally, it opens-source PyMARL, a deep MARL framework that includes state-of-the-art algorithms like QMIX, COMA, and IQL. The results demonstrate the performance of these algorithms on various SMAC scenarios, highlighting the need for more advanced coordination and exploration techniques in MARL.