1. A Machine Learning Driven Wireless IOT Sensor Network for Remote and Fragile Environments

Low-power processors, intelligent wireless networks, and low-power sensors coupled with Distributed Machine Learning (ML) running at the edge of the network, enables a multitude of sensors to be put anywhere: not just where communications and power infrastructure exists, but anywhere valuable information is gleaned. Pushing Machine Learning to the edge allows for immediate availability of predictive and intelligent classification and regression abilities, which is not dependent on available bandwidth/connectivity/latency of going back to the cloud, perform ML, and send results back to the edge. This is particularly important for a network designed to operate in a remote and fragile setting subject to unforeseen interruptions. The way these sensors are networked determines whether the sensors can provide reliability in face of disruptions. The platform/network being proposed has been architected with the following criteria in mind: Reliability Security Sensors can be put anywhere Ease of deployment Modularity Cost Our proposed solution is a Mobile Adhoc Network comprised of two types of nodes: Mesh Nodes, and Cellular Nodes. All nodes are communicating with each other at the physical layer using their own built-in Wifi, and at the link layer using the B.A.T.M.A.N.  (Better Approach to Mobile Adhoc Networking) advanced link layer mesh routing protocol. Link level routing (as opposed to Network level routing) uses raw ethernet packets to transfer routing data, making the network very resilient to unexpected interruptions in the network. This is because knowledge about routing is distributed throughout the network. Therefore, all nodes appear to be link local and are unaware of the network's topology as well as unaffected by any network changes. It is our intention to prove that optimization of this protocol will provide the best results for a resilient network in remote and fragile environments.   The choice of processor for the Mesh/Cellular Nodes were done carefully to be able to perform all necessary processing (for example Machine Learning predictive model execution), while maintaining low power, enabling all nodes to operate via solar power. It was proven that intelligent processing on the IoT processor enables utilization of cheaper & lower power sensors.   Various Machine Learning Algorithms were evaluated, and ultimately Random Forest was chosen for three reasons: Random Forest shows best performance Random Forest shows strongest robustness to loss of data Random Forest lends itself very well to a distributed implementation across many nodes We intend to show that a distributed implementation of the Random Forest algorithm at the application layer across the logistics network, coupled with the distributed routing intelligence of the BATMAN protocol, will enable the most agile and robust network for digital logistics needs, particularly in remote and fragile environments.