DescriptionWith rapid evolution of technology and growing use of wireless devices in our daily lives, mobile network is becoming one of the most promising platforms for many brand new applications. Several distinguished features make mobile network different from traditional computer networks, such as high mobility and unpredictable mobility patterns. With the networks becoming increasingly diverse and complex, it’s more and more difficult to know the properties of a network. Therefore, the need for ”learning” important network characteristics in such a dynamic knowledge setting becomes crucial. In this thesis, we show our research efforts to explore methods to learn important network properties, and improve the following three aspects of mobile networks: data management, load management, and identification services. The data management issue is most critical when there is no central infrastructure available or when the mobile-to-infrastructure communication bandwidth is limited. Since blindly uploading every piece of sensor data to a remote server is inefficient, local data aggregation is required to reduce the communication cost and improve efficiency. We propose the Geocache concept and the Boomerang anchoring protocol to address this issue, and further introduce adaptive learning methods to better deliver time-sensitive data. Our efforts in load management are focused on adaptive load-balancing schemes for wireless LANs where multiple access points are present. We propose a distributed access point selection scheme by which nodes select an appropriate access point to associate with, based on each individual devices channel utilization. This approach effectively reduces unnecessary reassociations and improves upper layer performance such as throughput and packet delivery delay. We further enhance the association protocol by using reinforcement learning to dynamically schedule the probing of neighboring access points (APs). By learning from past experience, we ultimately bring down the probing overhead. Lastly, we focus on the security aspect of the network by improving the identification process. We examine the problem of identifying different association protocols based on probing patterns, such as probing frequency and probing frame types. We apply learning methods to identify several association protocols and propose an approach which combines k-means clustering and Gaussian fitting to classify the association protocols.