Data-driven models used for predictive classification and regression tasks are commonly computed using floating point arithmetic preserving accuracy by automatic scaling even in high non-linear functions.With respect to distributed sensor networks like the IoT, sensor data is acquired on low-resource embedded systems and delivered to data servers characterized by big data volumes.In specific use cases and domains, local predictive modelling on low-power devices is desired or required.But heterogeneity of host platforms and dynamic programming disables machine code deployment.This work addresses Tiny ML on very lowresource devices (microcontrollers, less than 32 kB RAM and ROM) by using a stack-based Tiny Virtual Machine providing core ML operations to implement Decision Trees (DT), Artificial Neural Networks (ANN), and Convolutional Neural Networks (CNN).VM program code is always provided in textual format and compiled just-in-time to Bytecode to ensure portability, servicability, and mobility.Two damage diagnostics use-cases demonstrate the suitability of the VM approach, and even time consuming computational tasks do not compromise the overall responsiveness of the platform by using a real-time approach.This work addresses the underlying integer arithmetic operations required to implement efficient and fast computable ML models on microncontrollers.

Abstract4 Most Internet-of-Things (IoT) devices and smart sensors are connected via the Internet using IP communication directly accessed by a server that collect sensor information periodically or event-based.The spatial context (the environment in which the sensor or devices is situated) is not reflected accurately by Internet connectivity, and which is additionally not everywhere available.In this work, smart devices communicate connectionless and ad-hoc by using lowenergy Bluetooth broadcasting available in any smartphone and in most embedded computers.Bi-directional connectionless communication is established via the advertisements and scanning modes.The communication nodes can exchange data via functional tuples using a tuple space service on each node.Tuple space access is performed by simple evenat-based agents.The Bluetooth Low Energy Tuple Space (BeeTS) service enables opportunistic, ad-hoc and loosely coupled device communication with a spatial context.

Energy distribution as a main part of energy management in self-powered micro-scale networks like sensor networks is a challenge with the goal to satisfy a safe and reliable operational state on system and node level.Under the assumption that nodes are arranged in mesh-like networks with links posing the capability to transfer data and energy between nodes a selforganizing Mulit-agent System based on divide-and-conquer is deployed in this work successfully to distribute energy without a system/world level model and knowledge of single nodes about the system state.Different agent behaviour were investigated and the emergence evaluated.An exploring help strategy with energy deliver child agents showed the best and efficient overall behaviour.Mobile agents were programmed in JavaScript using the JavaScript Agent Platform that can be deployed in strong heterogeneous environments.

An agent processing platform suitable for distributed computing in sensor networks consisting of low-resource (e.g., material-integrated) nodes is presented, providing a unique distributed programming model and enhanced robustness of the entire heterogeneous environment in the presence of node, sensor, link, data processing, and communication failures. In this work multi-agent systems with mobile activity-based agents are used for sensor data processing in unreliable mesh-like networks of nodes, consisting of a single microchip with limited low computational resources. The agent behaviour, interaction, and mobility (between nodes) can be efficiently integrated on the microchip using a configurable pipelined multi-process architecture based on Petri-Nets. Additionally, software implementations and simulation models with equal functional behaviour can be derived from the same source model. Hardware and software platforms can be directly connected in heterogeneous networks. Agent interaction and communication is provided by a simple tuple-space database and signals providing remote inter-node level communication and interaction. A reconfiguration mechanism of the agent processing system offers activity graph changes at run-time.