The digital system of the future will face the growing challenge of controlling the system behavior in complex dynamically evolving environments. In this paper, we examine the applicability of a new management paradigm based on a reinforcement learning approach, where no preliminary specification of the system model is required. The learning agent identifies the most adequate control policies in live interaction with a partially observed system and provides it with autonomous management capabilities. We present the results of experimentation with cloud-based applications and discuss the technical challenges that need to be addressed in this field. Furthermore, we present the results of experimentation on a 5G network slice that hosts a cloud-based application in a multi-agent reinforcement learning setting, and demonstrate the value of information exchange between learning agents.
展开▼