A Multi-Agent Reinforcement Learning Architecture for Network Slicing Orchestration

The Network Slicing (NS) paradigm is one of the pillars of the future 5G networks and is gathering great attention from both industry and scientific communities. In a NS scenario, physical and virtual resources are partitioned among multiple logical networks, named slices, with specific characteristics. The challenge consists in finding efficient strategies to dynamically allocate the network resources among the different slices according to the user requirements. In this paper, we tackle the target problem by exploiting a Deep Reinforcement Learning approach. Our framework is based on a distributed architecture, where multiple agents cooperate towards a common goal. The agent training is carried out following the Advantage Actor Critic algorithm, which makes it possible to handle continuous action spaces. By means of extensive simulations, we show that our strategy yields better performance than an efficient empirical algorithm, while ensuring high adaptability to different scenarios without the need for additional training.