Distributed artificial intelligence for edge computing

In an increasingly wide range of work and personal life settings, sensors and micro-devices embedded into objects generate continuous data streams with high volume, velocity and heterogeneity. Such data can be analyzed to detect and infer knowledge about phenomenons and events of interest. By learning from data, Artificial Intelligence (AI) methods enable automating an ever larger amount of activities and decision-making tasks with comparable or better proficiency than human experts. Big Data applications based on pervasive Internet of Things (IoT) deployments are now a well-established reality: they feed large Machine Learning (ML) models, which are trained exploiting the huge computational resources of cloud computing infrastructures to offer increasingly accurate prediction capabilities on fresh data. However, the increasing miniaturization of IoT devices equipped with highly accurate sensors enables novel Cyber-Physical Systems (CPSs) with tight feedback loops coupling computation, communication and control tasks. CPS applications are expanding in sensitive fields like high-precision manufacturing, telemedicine, and self-driving vehicles. Those scenarios require real-time response, high computational and bandwidth efficiency, cost-effectiveness to support business scalability and strict data privacy constraints. For this reason, classical cloud-based approaches are progressively integrated with the Edge Computing (EC) architectural model, which distributes significant processing and storage resources at the edge of the local network, in closer proximity to field devices and sensors. This paradigm allows AI-based IoT applications to scale even more, as models are trained with massive amounts of data generated by large deployments of micro- and nano-devices, and ML inference achieves ever greater accuracy. In this context, the Edge Intelligence paradigm --which promotes the integration of EC and AI-- is increasingly adopted to execute inference on data at the border of local networks, employing models trained in the cloud. The next logical step in Cloud-Edge AI cooperation is to enable training tasks on edge nodes as well. However, as of now flexible approaches to combine Edge Intelligence with cloud infrastructures, allowing dynamic migration of training and inference tasks, are not available yet. This thesis proposes a Cloud-Edge AI microservice architecture, based on Osmotic Computing principles. A full pipeline consisting in data collection from local devices, preprocessing, AI model training and inferencing is supported either on the edge, on the cloud or both, exploiting computational resources opportunistically based on device status, available bandwidth, and application requirements on latency, prediction accuracy, and privacy. To demonstrate the feasibility of the proposed framework, a prototype has been realized with commodity hardware leveraging open-source software technologies. It has been used in a small-scale intelligent manufacturing case study, carrying out experiments on elapsed time and network activity in (i) data gathering, (ii) AI model training and validation, and (iii) prediction tasks. Obtained results validate the key benefits of the approach. In addition to architectural aspects, open issues still limiting the full applicability of EC include the heterogeneity of devices, services, and information that arise in pervasive contexts, the integrity of the gathered data, and the trustworthiness and dependability of autonomous decisions. The Semantic Web of Things (SWoT), coalescing the Semantic Web and IoT paradigms, has been proposed to overcome these problems. In SWoT environments, the dynamic exchange of knowledge fragments expressed in logic-based formalisms in volatile wireless networks of independent agents enables decentralized collaborative service discovery, autonomous decision and user decision support. A relevant problem in such scenarios consists in evaluating agreements and disagreements about knowledge produced by different interacting agents, in order to possibly reconcile conflicts and determine the best overall outcome to accomplish distributed coordination. In AI, Argumentation is recognized as a powerful formalism to negotiate and solve disagreements within a group of agents, which convey knowledge represented as a constellation of arguments and counterarguments. The argumentation literature provides a wealth of frameworks for agent decision-making and coordination. Nevertheless, few proposals leverage Semantic Web languages and technologies, which can provide a well-known formal model for arguments, well-studied inference algorithms for the assessment of argument relations and approaches to evaluate argument acceptability. This thesis presents a novel Bipolar Weighted Argumentation Framework, where arguments are modeled as Description Logics concept expressions in Web Ontology Language (OWL) 2, and their relations are assessed via semantic matchmaking, leveraging non-standard inference services with logic-based outcome explanation. Argument acceptability is computed via a novel propagation-based ranking semantics, which supports argument cycles and information fading. In order to make the proposal suitable for pervasive semantic agents in resource-constrained devices, optimizations in argument assessment and ranking evaluation are adopted, while a graph simplification approach via pruning is proposed and experimentally tested as a tunable trade-off between computational resource usage and accuracy of results. Validation of the approach has been carried out by means of a prototypical implementation of a player agent for the StarCraft II real-time strategy game, whose environment allows simulating the complexities of real CPS scenarios.