Jaguar: A Hierarchical Multi-Agent Deep Reinforcement Learning approach with Transfer Learning for StarCraft II
Deep Reinforcement Learning; StarCraft II; Multi-Agent; Transfer Learning; Hierarchical Architecture.
Real-time strategy games are environments that usually simulate real military situations and present a set of challenges for the field of Artificial Intelligence, such as the high complexity and large space of actions and states, partially observable maps and the fact that they deal with multiple agents at the same time, in addition to the tasks being able to be performed within the scope of micromanagement or macromanagement. In particular, Reinforcement Learning has been highlighted in the application and evolution of techniques capable to deal with these challenges, especially with the advent of Deep Reinforcement Learning. A systematic literature review was conducted with the goal of understanding and summarizing which environments, techniques, tools and architectures make up the state of the art of Deep Reinforcement Learning in Real-Time Strategy games. From the information obtained by the review, this work proposes to develop a Hierarchical Multi-Agent approach focused on the game StarCraft II, using Transfer Learning and action masking techniques to aid the agent training consume less resources and obtain satisfactory results even for more complex scenarios.