"Know Thyself" - Computational Self-Reflection in Intelligent Technical Systems

Self-improving system integration: Mastering continuous change

“…More specifically, the paper delineates self-integration as an "ongoing autonomous process for linking a potentially large set of heterogeneous computing systems, devices, and software applications; to meet system goals", and self-improvement as "inventing solutions where none are available and optimising the selection of solutions over time" (which we took as basis for Section 2. Conceptually related are Computational Self-Awareness [38,39] and Self-Reflection [40]. This is the ability of a system to learn about itself on an ongoing basis, in order to reason about and improve its own behaviour in a complex environment.…”

“…b) Problem classes: The application scenarios underlying the various theoretical considerations range from autonomous vehicles [42] over vehicular traffic control [45] and route guidance [46], robotics and simulation [40], emergency response [47], to intelligent surveillance systems [44] and energy systems [48]. From these contributions, together with those that take a more generalist system-of-systems stance [6,43,37] or that involve specific interaction models, e.g., [41], we can extract the following core problem classes:…”

“…Assessment and research gap: While these problem classes are indicative, many of their details remain to be examined for their formal understanding and full applicability across domains. The layered system model proposed in [40] constitutes a step in that direction, differentiating between the abstract layers Reaction, Adaptation, Reflection and Collaboration in a heterogeneous set of scenarios. Similarly, [42] identifies the following central challenges that need to be tackled before successful self-integration becomes truly viable: resolution for concepts, temporal planning, trackable events, and merging.…”

“…The agent is not aware of its performance and state per se but has its goals ingrained and operates using a utility function. This utility function allows the system to make decisions based on its most recent actions and current state: They identify and assess their current state, their behaviour, and their most recent actions [40,47]. The "Wrappings" approach [113] which was discussed previously has specifically been designed for achieving this task.…”

Lifelong Learning and Collaboration of Smart Technical Systems in Open-Ended Environments -- Opportunistic Collaborative Interactive Learning

“…Altogether, this new generation of smart systems will exhibit various self- * properties (e.g., self-and context-awareness or self-reflection [5]) and active learning [6] will play a key role to build such systems. However, the active learning techniques needed here are far beyond the currently available techniques with their limitations regarding number and availability of information sources (typically one that is omnipresent), quality of information sources (typically omniscient), cost for queries (mostly not considered), type of queries (typically only one: labels for data objects), assessment of own knowledge (mostly not regarded), etc.…”