This project made fundamental contributions toward making multi-layered communication and computing infrastructures capable of autonomously deciding how to acquire, transport and process information to accomplish a global task. Such a goal can only be achieved by introducing algorithms across the logical layers of the system that can adapt the overall information management strategy based on past experience and current state. The project attacked different aspects of the problem. First, we built a novel class of collaborative and adaptive classification algorithms for edge computing. Using a stochastic control loop between a mobile device and the edge server, our algorithms can autonomously adapt the parameters of a simple classifier deployed at a mobile device to follow the semantic state of the data, expressed as the sample distribution. The approach we developed allows computationally weak layers of the system to perform analysis and filtering of complex data. 

Within the project, an experimental platform was built that included multiple interconnected autonomous Unmanned Aerial Vehicles (UAV) and ground edge servers. We developed a middleware allowing experience-based control of how information is processed to meet data analysis delay constraints imposed by real-time applications controlling the navigation of the UAVs based on visual input. Within this context, we developed the concept of distributed redundant task allocation, where the distribution of the computing tasks to one or multiple servers is determined at  fine temporal granularity by deep reinforcement learning agents deployed at the UAV. Decision making is based on a broad definition of observable state, where state variables span across application, network and telemetry blocks. The middleware, datasets and models were all released as open source tools.

In order to facilitate the offloading of complex data analysis algorithms, in this project we introduced the concept of split Deep Neural Networks (DNN) where the architecture of the models is modified to embed an encoder-decoder structure realizing a supervised form of in-model compression. Specialized DNN training approaches were developed to maximize the compression gain, and thus minimize system load and communication time, while resulting in minimal degradation of task performance. The models and code were released as open source tools to the research community.

The project also made several technical contributions to sensor placement and target tracking by developing new algorithms based on dynamic active average consensus and distributed submodular maximization. These results allow efficient and effective distributed deployments for a group of mobile agents providing a service for a dense set of targets with limited knowledge.

The project supported workforce training at all levels, including Ph.D., master and undergraduate students from minorities. The interdisciplinary nature of the project attracted considerable attention from undergraduate students, who were involved in many research tasks. We expect that these activities will enhance the curriculum of the students and their probability of being admitted to prestigious master and Ph.D. programs. We estimated that more than 30 undergraduate students were involved in project-related tasks. The project resulted in many impactful publications across different communities.

Last Modified: 05/27/2022
Modified by: Marco Levorato