Abstract : The recent evolutions of the Internet have been materialized by the emergence of new distributed application as well as the rapid increase of the number of network technologies (wireless, mobile. . . ). These advances have allowed new services to be proposed by operators using a new type of terminals (laptop, PDA. . . ). The major socio economic stake that these advances represent is materialized by the future ambient Internet, which is both ubiquitous and intelligent, through which the user will be able, whatever his location or access point, benefit from the best possible Quality of Service (QoS) given its current applicative and network environment. In this context, the thesis presents an architecture as well as models and algorithms that allow for a dynamic self-adaptive composition of services provided by existing QoS mechanisms. The approach is based on the dynamic and coordinated adaptation of both the behaviour and the architecture of the protocols that compose the communication stack. The work finds its inspiration in the theory of artificial intelligence and learning and provides the specification, implementation and evaluation of a self-adaptive communication system taking into account both, the applicative requirements for the transported streams and the constraints of the environment on which communication takes place. Finally, the evaluation of the decision and learning models illustrates how the system allows to fulfil its objective and validates the concepts that are proposed in this thesis.