Wireless Networks operating in ISM bands, with Wireless Local Area Networks (WLANs) as their most representative technology, are nowadays the most common wireless Internet access technology. A new generation of WLANs, the one that will be part of the 5G ecosystem, is already under development and will make its appearance in the forthcoming years. However, although this new generation of WLANs will be technologically very advanced, it will still be based on a traditional conception, where the WLAN is a simple standalone wireless extension of the fixed network, and operate following a rigid approach, with all functionalities pre-configured. Clearly, that situation will not be able to satisfy the demanding users’ requirements in the expected dense, highly dynamic, heterogeneous and complex wireless network future scenarios, where bandwidth-hungry tactile-like real- time multimedia interaction with people and things, real and virtual, will require a high degree of flexibility from the network to dynamically adapt to many different situations. In addition, most of the time those networks are operating in a never ending transient state, in which all expected performance figures computed using traditional optimization models are completely useless.

The challenge is to provide a feasible solution to maximize the wireless networks performance in such complex and dynamic scenarios.

To solve that challenge, a possible solution is that each individual network makes decisions regarding the use of the channel resources based on its own observations of the spectrum occupancy, and the gathered experience from the result of its previous actions. Those decisions may be simple, such as switching to a different channel, increase / decrease the transmission power, the channel bandwidth, or the energy detection level. In case assistance from the cloud is provided, more complex decisions can be made, including those that affect one or more networks such as distributing the users efficiently between different access points. To do that, we need to provide to the group of wireless networks with the means to extract knowledge from the environment, and make decisions based on it that result in a better performance. However, there is always an associated overhead and complexity when adding those new functionalities which requires to be carefully considered.

Therefore, the main goal of this MdM project is to show that combining machine learning techniques to allow networks learn what are the best action to take in every situation, and data analytic techniques to study large data-sets containing miscellaneous information to make global recommendations, wireless networks can be able to adapt to many different situations, achieve an efficient use of the resources, and operate close to the optimal point. Namely, we want to provide answers to the following questions: Where, how and when can we use machine learning techniques to optimize the operation of wireless networks? Can we extract useful information from analyzing collected data from the networks and users in order to recommend new network configurations? Can we enforce collaboration strategies to share information between networks over the radio interface or through the cloud? What is the overhead (or downside) of using those techniques? Overall, is it possible to evaluate how much close those techniques are able to make the networks operate close to the desired point?