Reinforcement Learning Meets Cognitive Situation Management: A Review of Recent Learning Approaches from the Cognitive Situation Management Perspective

摘要

With Reinforcement Learning (RL), artificial agents learn reaching their goals “in the wild”, i.e., from interacting with their environments. By learning to perform the correct action(s) in the given situation, RL thus adopts an action or decision- centric problem orientation. Conversely, the field of Cognitive Situation Management (CogSiMa), more originating from the control field, focuses on managing the encountered situations, i.e., environment states, such that the desired goal situations are reached or maintained. Whereas both fields of research thus appear complementary in pursuing similar overall goals, RL and CogSiMa have largely evolved independently from each other, leading to terminological gaps, misconceptions and unawareness of potentially related research. The present review attempts to bridge these gaps by providing an integrated framework highlighting the intersections between RL and CogSiMa: We outline how RL in real-world problem domains relates to CogSiMa, aim to bridge the terminological gaps between these distinct communities, and hope to provide the grounding for a cross-fertilization between these distinct research areas. We contribute a review of recent RL developments and discuss their implications and potential for CogSiMa.