Open Access

Anthony Stein

• Self-adaptive systems are increasingly endowed with Artificial Intelligence technology in order to enhance system autonomy. Most prominently, algorithms from the research field of Machine Learning are utilized to allow autonomous agents to continually strive for increasing the system's utility based on experiences made over time. The resulting self-learning adaptive systems are typically deployed in dynamic non-deterministic environments which are characterized by continuing change and stochasticity regarding condition observations and utility measurements. This often leads to circumstances where the inner learning mechanisms are exposed to so far unseen and unanticipated system states for which they lack sufficient knowledge about how to react appropriately. In this thesis a novel, technical notion of Knowledge Gaps in self-learning adaptive systems is developed in order to characterize exactly this challenge. Knowledge gaps are assumed to be existent within the continuallySelf-adaptive systems are increasingly endowed with Artificial Intelligence technology in order to enhance system autonomy. Most prominently, algorithms from the research field of Machine Learning are utilized to allow autonomous agents to continually strive for increasing the system's utility based on experiences made over time. The resulting self-learning adaptive systems are typically deployed in dynamic non-deterministic environments which are characterized by continuing change and stochasticity regarding condition observations and utility measurements. This often leads to circumstances where the inner learning mechanisms are exposed to so far unseen and unanticipated system states for which they lack sufficient knowledge about how to react appropriately. In this thesis a novel, technical notion of Knowledge Gaps in self-learning adaptive systems is developed in order to characterize exactly this challenge. Knowledge gaps are assumed to be existent within the continually growing but limited knowledge bases of these systems. This gap-centric perspective is transferred to the acquisition process of incrementally knowledge building systems. Accordingly, it is intended to (1) pave the way for the development of novel techniques which aim at countering such knowledge gaps, and, (2) to strengthen the self-reflective capability of self-learning adaptive systems with regard to their current knowledge. The former aspect constitutes the main topic of this thesis. On the basis of a well-known Evolutionary Rule-based Machine Learning technique, which has been applied several times in the context of Organic Computing research, the algorithmic structure of this type of algorithms is enhanced toward explicitly counter existing gaps in their incrementally evolving knowledge bases. It will be demonstrated how the exploitation of so far acquired knowledge elements can be improved by further incorporating raw experiences in a transductive manner. Transduction here means to immediately leverage already made and remembered experiences instead of inducing a model first and deducing to new situations afterward. Furthermore, the initialization of newly constructed knowledge elements is enhanced. Again, it will be explained how to make transductive use of already existing but not directly matching knowledge elements in the proximity of the currently queried problem space niche about which the algorithm is not confident yet. This knowledge transduction is realized by utilizing methods from the domain of scattered data interpolation within the XCS Classifier System -- the most prominent representative from the class of Michigan-style evolutionary rule-based machine learning systems. Results on a range of conducted empirical validation studies are reported to corroborate the hypothesized benefits of transductive knowledge inference in XCS by means of interpolation. Additionally, plausibly applicable parts of the introduced methodologies are transferred to a realistic application scenario in the context of self-adaptive traffic light control. In order to capture the aforementioned second advantage of knowledge gap-centric learning -- the advance of the self-reflection property of self-learning adaptive systems -- a novel research direction for autonomous learning which is termed Proactive Knowledge Construction is proposed and first steps toward Proactive Learning Classifier Systems are taken. It is elaborated on how concepts from the domain of Active Learning can be incorporated within the Organic Computing Multi-layer Observer/Controller reference architecture in order to actively seek and bridge knowledge gaps within these systems. Furthermore, in order to substantiate the rationale behind the general concept of proactive knowledge construction, an initial formal proof is outlined and a first methodology to implement the envisioned proactive behavior is delineated in the last part of this thesis.…