Dynamic Traffic Control Clustering

Anticipatory optimal network control is defined as the problem of determining the set of control actions that minimizes a network-wide objective function, while not only taking into account local consequences on the propagation of flows, but also the global network-wide routing behavior of the users. Such an objective function is, in general, defined in a centralized setting, as knowledge regarding the whole network is needed to correctly compute it. Reaching a sufficient level of centralization is, however, unrealistic, as multiple authorities are influencing different portions the network, separated either hierarchically or geographically. This is the motivation for this work, in which the authors introduce a decomposition mechanism for the global anticipatory network traffic control problem based on dynamic clustering. Rather than performing full decomposition of the problem, this technique allows recognizing when and which controllers should be grouped in clusters, and when, instead, they can be optimized separately. The practical relevance with respect to their motivation is that it allows identification of those network traffic conditions in which multiple actors need to actively coordinate their actions, or when unilateral action suffices for still approximating global optimality. This clustering procedure is based on well-known algebraic and statistical tools, who exploit the network’s sensitivity to control and its structure to deduce coupling behaviour. The authors perform several tests in order to assess their newly introduced procedure’s performances, in comparison with fully decomposed and fully centralized anticipatory optimal network control, and show that their approach is able to outperform both centralized and decomposed procedures.