A Generic Architecture for Restricted On-line Learning of Network Protocol Parameters

The Organic Network Control (ONC) system has been developed to adapt the parameters of network protocols dynamically to changes in the environment in order to increase the protocol´s performance. In order to apply the system to new environments without the need of providing knowledge it is equipped with a simulation-based rule-generation and evaluation component on top of an on-line rule selection mechansim. When aiming for a generic system that is capable of controlling varying network protocols from all layers of the protocol stack, a simulation-based approach has some drawbacks. This paper describes the limitations of the current approach and discusses possibilities to cope with these problems. As a result a more general architecture for restricted on-line learning is presented. Based on the definition of the resulting research questions, an outlook on the upcoming work on ONC is given.