Open Access

Stefan Rudolph

• Since the 1990s, we see an incremental miniaturization of computers, and, consequently, a computerization of everyday objects. This trend has strongly increased over the last decades and there is no sign of a reversal. The Internet of things (IoT) and cyber-physical systems (CPS) are current domains, which are a direct continuation of this development. Most of the time it is not possible to model such a high number of devices at design-time. Therefore, Organic Computing aims to overcome the issues that appear with the control of systems with high complexity by implementing the concepts of self-adaption and self-organization in technical systems, i.e., enable the systems to adapt and manage themselves during runtime. However, due to the complexity and size of modern systems, it is often not clear to whom the systems should adapt to or organize with. Such complexity appears because of high numbers of autonomous systems, which can bear a high degree of complexity in itself, but even moreSince the 1990s, we see an incremental miniaturization of computers, and, consequently, a computerization of everyday objects. This trend has strongly increased over the last decades and there is no sign of a reversal. The Internet of things (IoT) and cyber-physical systems (CPS) are current domains, which are a direct continuation of this development. Most of the time it is not possible to model such a high number of devices at design-time. Therefore, Organic Computing aims to overcome the issues that appear with the control of systems with high complexity by implementing the concepts of self-adaption and self-organization in technical systems, i.e., enable the systems to adapt and manage themselves during runtime. However, due to the complexity and size of modern systems, it is often not clear to whom the systems should adapt to or organize with. Such complexity appears because of high numbers of autonomous systems, which can bear a high degree of complexity in itself, but even more as a result of the non-trivial interactions among these systems. This leads to mutual influences which can be either direct, i.e., easily observable, or indirect, i.e., hard to observe or hidden. Not easily observable influences arise especially because of indirect couplings or influential paths that are not perceptible by the individual systems, e.g., because the environment reacts to several autonomous systems operating on it at the same time. To address these mutual influences, it is necessary to uncover them during runtime and enable the systems to self-adapt to them. Therefore, the basic theme of this thesis is a computational approach to acquire knowledge about mutual influences among systems or system parts that allow them to be controlled more efficient by enabling interaction and averting to interfere with each other. Such a methodology should allow for backwards compatibility as well as applicability in heterogeneous systems and utilization for autonomous application during the runtime of a system. In this thesis, the algorithmic approach to influence detection and several aspects of its application are covered. The proposed method is based on stochastic dependency measures that are applied to quantify how important the other systems are for the outcome of the system itself based on past experience, i.e., the systems measure the correlation between their reward signal and the configuration of other systems. Especially, the following dependency measures are evaluated for this purpose: the Pearson correlation, the Kendall rank correlation, the Spearman rank correlation, the distance covariance, the mutual information, and the maximal information coefficient. The basic methodology is expanded in four ways: (i) the incorporation of mutual reactions between several components in the system by conditioning the measurement of the influence by these other components, (ii) the detection of delayed influences, i.e., influences that do not manifest immediately but after a time period, by measuring the reward signal against past configurations of other systems, (iii) the application at runtime by relying on partially randomized configurations, and (iv) the adaption of the systems to the influences on each other by employing reinforcement learning algorithms. The approaches are evaluated on elementary use cases and two real-world applications. The latter are smart camera networks from the IoT domain and smart factories as a representative of CPS. The results show that the method can detect all influences on each of the applications. Eventually, practical advice on the application of influence detection in different system classes are given.…