This thesis investigates learning-based resource allocation techniques for future wireless networks (FWNs). Motivated by recent technological developments, two types of FWNs are studied: energy harvesting (EH) wireless sensor networks (WSNs), and high-capacity cellular networks (HC-CNs) with caching capabilities. In an EH-WSN, each node is powered by a rechargeable battery and harvests energy from the environment. First, a multi-access throughput optimisation problem is studied, when the access point schedules EH sensor nodes without the knowledge of their battery states. A low-complexity policy is shown to be optimal in certain cases, and a scheduling algorithm, which takes into account the random processes governing the energy arrivals in the system, is proposed, and compared to an upper bound. Second, a point-to-point communication system with an EH transmitter is considered. Since the characteristics of the environment in which the sensor will be deployed are not known in advance, we assume no a priori knowledge of the random processes governing the system, and propose a learning theoretic optimisation for the system operation. The performance of the proposed algorithm is compared to that of two upper bounds, obtained by providing more information to the transmitter about the random processes governing the system. We then turn our attention to content-level selective offloading to an infostation terminal in an HC-CN. The infostation, which stores high data-rate content in its cache memory, allows cellular users in the vicinity to directly download the stored content through a broadband connection, reducing the latency and the load on the cellular network. The goal of the infostation cache controller is to store the most popular content such that the maximum amount of traffic is o oaded to the infostation. The optimal cache content management problem when content popularity is unknown is studied, and a number of algorithms to learn the content popularity pro le are proposed. The performances of these algorithms are compared to that of an informed upper bound, obtained when the content popularity pro le is known.