Data transmission over IP networks is, by the very nature of this protocol, based on best effort, with all flows being treated identically as a principle. Any of the existing Traffic Engineering (ET) alternatives that enable quality of service (QoS) in IP networks (such as MPLS, IntServ or DiffServ), by themselves do not implement sufficiently comprehensive solutions. A suitable adaptation function can be implemented to help in the operation of networks with or without these ET models.

Bandwidth provision and other operational parameters in a network are complex to calculate due to dynamic traffic patterns. As bandwidth reservation models do not scale well, a learning approach may be more suitable for this problem. In particular, algorithms Reinforcement Learning (RL) allow the implementation of distributed decisions, taken by agents in segments of the periphery of the network. If these segments are connected to the same core, the network can improve its ability to respond to situations for which it was not specifically configured, and the quality of service can be less affected when there are potential impairments.

A RL model distributed by the Internet of Things things, to adapt to the specific conditions of the network, can be beneficial for the system as a whole, without presenting excessive signaling load, single point of failure or too long time for taking the actions. correct decisions.

speakers

Laercio Cruvinel

Department of Science and Technologies