IETF's Differentiated Services (DiffServ) is a scalable, distributed architecture aimed to provide a fair distribution of network resources according to expected levels of service. However, variable traffic and performance expectations may require an adjustment of management policies so that congestion situations are avoided and service parameters are met. Existing solutions that improve the standard DiffServ architecture do not cope well with this dynamics and network resources are not efficiently managed, preventing user expectations from being properly fulfilled. In this paper we show how a reinforcement learning approach can optimize the choice of adaptation profiles for the optimal adjustment of ingress policies when a pre- congestion state is detected. Numerical results from simulations showed improvements for priority traffic (video) without impacting excessively the other traffic.

Quality of service for video and audio transmissions over IP is bound by the best-effort nature of this protocol. The road for achieving optimal behavior for selected flows of traffic includes better controlling and tuning one or more elements of the transmission - the characteristics of the traffic itself or the supporting hardware and software. This tuning may be static or dynamic, profile-based or adaptive. This paper presents results and insights of using an architecture for adapting QoS parameters in a DiffServ-enabled network, in an effort to dynamically reach the best choice of values in each given situation. The architecture is named distributed dynamic quality of service - DDQoS - and includes separate, but interoperating, models for the core and for the edge of the network. The momentum experienced by the transmission of multimedia content over the Internet, and the extended range of options for adapting this kind of traffic, derived from its particular characteristics when compared with other data, motivate this work on adaptation mechanisms for improving the quality of service.