NeTS: Small: Learning-Guided Network Resource Allocation: A Closed-Loop Approach

基于网络测量和用户行为数据，最近的许多工作研究了使用机器学习技术对网络效用和用户体验的建模和预测。虽然它提供了重要的见解，但预测本身往往不是网络的最终目标。理想情况下，网络可以识别体验不佳的用户，并采取适当的措施来主动改善整体性能。为了实现这一目标，该项目倡导一种闭环方法，该方法使用学习辅助效用模型来明确指导网络中的资源分配，并使用反馈来验证和改进学习的效用模型。这项研究为理解、设计和分析学习模型辅助的资源优化算法提供了重要的见解。此外，由于其通用性，该闭环方法可以应用于具有以下特征的其他系统：1）系统太复杂，不能仅依赖于领域知识来构建白盒效用模型; 2）存在足够的数据，使得可以学习效用模型; 3）为了最大化总体效用，可以对影响效用值的某些控制变量进行优化;以及4）存在反馈回路，使得可以观察控制的效果。该项目的成果可应用于不同学科的此类系统。由于网络效用函数的未知性和噪声性，以及在高维、耦合资源约束和非凸优化的背景下，利用所提出的框架是非常具有挑战性的。为了应对这些挑战，该项目考虑了两种互补的方法：贪婪方法和综合方法。贪婪方法在应用不同的学习模型方面具有很大的灵活性，在实践中可以更好地适应不同的应用场景，但很难分析。集成的方法建立在高斯过程（GP）的土匪，集成了构建的模型和模型的不确定性，在资源分配决策。这种方法更适合理论分析，尽管具有很大的挑战性。在这两种方法中，都需要根据学习的模型来优化资源分配。该项目的贡献来自于解决相应的非凸优化问题。最后一步是使用闭环反馈来构建更好或最优的效用模型。综合方法的目的是开发层次GP强盗算法降维，理想的理论性能保证。贪婪的方法利用一般学习模型的扰动探索方案，并力求实用性和通用性。

Based on network measurement and user behavior data, much recent work has studied the modeling and prediction of network utility and user experience using machine learning techniques. While it provides important insights, prediction itself is often not the ultimate goal in networks. Ideally, a network could identify users with poor experience and take proper actions to proactively improve the overall performance. To achieve this goal, the project advocates a closed-loop approach that uses learning-aided utility model to explicitly guide resource allocation in networks and uses feedback to (in)validate and improve the learned utility model. This investigation provides important insights in understanding, designing, and analyzing learning-model-aided resource optimization algorithms. Furthermore, because of its generality, this closed-loop approach can be applied in other systems with the following characteristics: 1) the system is too complex to rely on domain knowledge only to build a white-box utility model; 2) there exists sufficient data so that a utility model can be learned; 3) to maximize the overall utility, one can optimize over certain control variables that affect the utility value; and 4) there exists a feedback loop so that the effect of the control can be observed. The outcome of the project can be applied to such systems in different disciplines. Utilizing this proposed framework is highly challenging due to the unknown and noisy nature of the network utility function, and in the context of high dimensionality, coupled resource constraints, and non-convex optimization. To address these challenges, the project considers two complementary approaches: a greedy approach and an integrated approach. The greedy approach has much flexibility in applying diverse learning models, which may fit different application scenarios better in practice, but is difficult to analyze. The integrated approach builds upon Gaussian Process (GP) bandits that integrate both the constructed model and model uncertainty in resource allocation decisions. This approach is more amenable to theoretical analysis, although highly challenging. In both approaches, one needs to optimally allocate resource based on the learned models. The contribution of the project comes from solving the corresponding non-convex optimization problems. The last step is to use the closed-loop feedback to build a better or optimal utility model. The integrated approach aims to develop hierarchical GP bandit algorithms for dimensionality reduction, ideally with theoretical performance guarantees. The greedy approach leverages perturbed-exploration schemes for general learning models and strives for practicality and generality.