Resource sharing in network slicing and human-machine interactions

In this thesis we explore two novel resource allocation models. The first addresses challenges associated with dynamic sharing of network resources by multiple tenants/services via network slicing. The second focuses on a data-driven approach to the optimization of resource allocation in interactive human-machine processes. In our first thrust we investigate how to allocate shared storage, computation, and/or connectivity resources distributed amongst multiple tenants/ virtual service providers which have dynamic loads. It is expected that next generation of wireless network will be shared by an increasing number of data-intensive mobile applications (e.g., autonomous cars, IoT, interactive 360° video streaming), and tenants/service providers. A key functional requirement for such infrastructure is enabling efficient sharing of heterogeneous resource among tenants/service providers supporting spatially varying and dynamic user demands, both from the point of view of enabling the deployment and performance management to diverse service providers and/or tenants, as well as means to increase utilization and reduce CAPEX/OPEX associated with deploying possible new infrastructures. To that end, we propose a novel dynamic resource sharing policy, namely, Share Constrained Proportional Fair (SCPF), which allocates a predefined ‘share’ of a pool of (distributed) resources to each slice. We provide a characterization of the achievable performance gains over General Processor Sharing (GPS), and Static Slicing (SS), i.e., fixed allocation of resources to slices. We also characterize the associated share dimensioning problem, asking when a particular set of load profiles and QoS requirements are feasible, as well as what should be an appropriate pricing strategy. We further consider possible slice-based admission control scheme where slices engage in an underlying game to maximize their carried loads subject to performance requirements. In order to accommodate settings where one would wish to provision different types of resources which are coupled through user demands, we generalize SCPF to a more general resource allocation criterion, namely, Share Constrained Slicing (SCS), which extends traditional α—fairness criterion, by striking a balance among inter- and intra-slice fairness vs. overall efficiency. We show that SCS has several desirable properties including slice-level protection, envyfreeness, and load-driven elasticity. In practice, mobile users' dynamics could make the cost of implementing SCS high, so we also study the feasibility of using a dynamically weighted max-min fair policy as a surrogate resource allocation scheme. For a setting with stochastic loads and elastic user requirements, we model the user dynamics under SCS as a queuing network and establish the stability condition. Finally, and perhaps surprisingly, we show via extensive simulation that while SCS (and/or the surrogate weighted max-min allocation) provides inter-slice protection, they can also achieve improved job delay and/or perceived throughput, as compared to other weighted max-min based allocation schemes whose intra-slice weight allocation is not share-constrained, e.g., traditional max-min and/or discriminatory processor sharing. In our second thrust we study how to optimize resource allocation in the context of human-machine interactions. Examples of such processes could include systems aimed at assisting humans in interactive learning, workload allocation, or web-search advertising. We devise an innovative framework to enable the optimization of a reward over an interactive process in a data-driven manner. This is a challenging problem for several reasons: (1) humans' behavior is not easily modeled and may reflect biases, memory and be sensitive to sequencing, all of which should/could be inferred from data; (2) because these interactions are typically sequential and transient, inferring such complex models for human behavior is difficult; (3) furthermore, in order to collect data on human-machine interactions one must choose a machine policy which in turn may bias inferences on human behavior. In this thesis we approach the problem of jointly estimating human behavior and optimizing machine policies via Alternating Entropy-Reward Ascent (AREA) algorithm. We characterize AREA in terms of its space and time complexity and convergence. We also provide an initial validation based on synthetic data generated by an established noisy nonlinear model for human decision-making