On Behaviour-Based Network Inference and Distribution-Based Network Control

Abstract

Complex networks are prevalent structures throughout technological systems, and are also used to model many non-technological systems as well. Application domains that make use of networks range from financial systems, to biology/medicine, to online social networks and agent-based systems. There is a strong desire to control these types of systems, to avoid catastrophic failures, increase system stability, or achieve some known system goal. The development of automated controllers for these types of systems is a complex problem that involves several key subproblems, including the selection of a control node set and the generation of control signals to be injected into the network. Previous research involving network control has typically assumed the underlying network connections are precisely known and has also taken a strictly vector-based view of system state. This thesis expands on the existing network control work in two significant directions. First, this thesis investigates algorithmic, behaviour-based methods for predicting links within networks. This involves using transfer entropy measurements, calculated between time series of actions generated by participating agents. This approach could be used to predict underlying networks in unobservable problem domains (e.g., financial systems) or to identify links that are truly influential within observable problem domains (e.g., online social networks). A number of prediction algorithms are proposed and compared, several of which attain high levels of accuracy, even with a limited amount of available system information. Second, this thesis eschews the traditional vector-based view of system state within network control problems, proposing a novel, distribution-based approach. One of the most studied control goals in the existing research has involved moving a system between two vector-based states. Distribution-based control, however, identifies state distributions as targets for control, which is arguably a more expressive and suitable approach for many problem domains. The effect of various network parameters on control success is investigated within a distribution-based control problem, with microscopic analysis of subset distributions being used to demonstrate why control is more difficult in certain scenarios. This information is used to inform the creation of new control node selection algorithms, with statistically significant improvements over the highest rated heuristic from previous research being realized.