In a scenario where multiple wireless networks of different technologies are available, this work addresses the problem of the design of a cognitive engine, core of a cognitive radio device, able to perform the surrounding radio environment recognition and the network selection with the final goal of maximization of final user Quality of Experience (QoE). Particular focus is put on the requirement of simplicity of all the elements involved, from hardware to algorithms, in order to keep in mind the importance of its practical realizability.Two aspects were investigated. For the surrounding radio environment recognition step, a network identification and automatic classification method based on MAC layer features was proposed and tested. As regards the network selection, Key Performance Indicators (KPIs), i.e. application layer parameters, were considered in order to obtain the desired goal of QoE. A general model for network selection was proposed and tested for different traffic types, both with simulations and a practical realization of a demonstrator (implemented as an application for Android OS). Moreover, as a consequence of the originated problem of when measuring to estimate a network performance and when effectively using the network for data transmission and reception purposes, the multi-armed bandit problem (MAB) was applied to this context and a new MAB model was proposed, in order to better fit the considered real cases scenarios. The impact of the new model, that introduces the distinction of two different actions, to measure and to use, was tested through simulations using algorithms already available in literature and two specifically designed algorithms.