A learning classifier system for emergent team behavior in real-time POMDP

Often the only solution for many complex and dynamic real-world situations is a crucial concurrent cooperation and coordination divided into tasks and subtasks, i.e. team behavior [1]. This research focus on such problems under real-time constraints, distributed control and decentralized knowledge. Existent frameworks and simulation systems were designed relying heavily on a priori knowledge of experts and introducing little or nothing of Machine Learning (ML). Therefore, the goal here is to develop a team of agents inspired by team behavior as found in Nature - emergent and adaptive - applying only ML on the action-selection decision process. Such team would reduce time and resources in the design of autonomous teamwork while keeping equivalent performance in comparison to a heuristic-based approach. Applying unbiased methods and a divide and conquer strategy, we achieved individual actions that emerge into the aimed collective behavior, not once requiring plans, common beliefs or agreed intentions.