http://www.ronpub.com/publications/search.php?journal=ALL&author=Mauricio+Fadel+Argerich&exactauthor=on&title=&abstract=&volume=&issue=&year1=&year2=&searchtype=advanced This feed contains the result of an user-defined search in RonPub publications en-us https://www.ronpub.com/ojiot/OJIOT_2018v4i1n06_Fuerst.html http://nbn-resolving.de/urn:nbn:de:101:1-2018080519303887853107 Traditional device-cloud architectures are not scalable to the size of future IoT deployments. While edge and fog-computing principles seem like a tangible solution, they increase the programming effort of IoT systems, do not provide the same elasticity guarantees as the cloud and are of much greater hardware heterogeneity. Future IoT applications will be highly distributed and place their computational tasks on any combination of end-devices (sensor nodes, smartphones, drones), edge and cloud resources in order to achieve their application goals. These complex distributed systems require a programming model that allows developers to implement their applications in a simple way (i.e., focus on the application logic) and an execution framework that runs these applications resiliently with a high resource efficiency, while maximizing application utility. Towards such distributed execution runtime, we propose Nandu, an actor based system that adapts and migrates tasks dynamically using developer provided hints as seed information. Nandu allows developers to focus on sequential application logic and transforms their application into distributed, adaptive actors. The resulting actors support fine-grained entry points for the execution environment. These entry points allow local schedulers to adapt actors seamlessly to the current context, while optimizing the overall application utility according to developer provided requirements. https://www.ronpub.com/ojiot/OJIOT_2020v6i1n04_Fuerst.html https://www.ronpub.com/ojiot/OJIOT_2020v6i1n04_Fuerst.html Machine learning (ML) applications in Internet of Things (IoT) scenarios face the issue that supervision signals, such as labeled data, are scarce and expensive to obtain. For example, it often requires a human to manually label events in a data stream by observing the same events in the real world. In addition, the performance of trained models usually depends on a specific context: (1) location, (2) time and (3) data quality. This context is not static in reality, making it hard to achieve robust and transferable machine learning for IoT systems in practice. In this paper, we address these challenges with an envisioned method that we name Knowledge Infusion. First, we present two past case studies in which we combined external knowledge with traditional data-driven machine learning in IoT scenarios to ease the supervision effort: (1) a weak-supervision approach for the IoT domain to auto-generate labels based on external knowledge (e.g., domain knowledge) encoded in simple labeling functions. Our evaluation for transport mode classification achieves a micro-F1 score of 80.2%, with only seven labeling functions, on par with a fully supervised model that relies on hand-labeled data. (2) We introduce guiding functions to Reinforcement Learning (RL) to guide the agents' decisions and experience. In initial experiments, our guided reinforcement learning achieves more than three times higher reward in the beginning of its training than an agent with no external knowledge. We use the lessons learned from these experiences to develop our vision of knowledge infusion. In knowledge infusion, we aim to automate the inclusion of knowledge from existing knowledge bases and domain experts to combine it with traditional data-driven machine learning techniques during setup/training phase, but also during the execution phase.