Abstract |
In future tactical environments, situational awareness will require fusion of data from sensors distributed across the battle domain, while computing resource access and network bandwidth are likely to be limited. We consider this problem as the ultimate limit of edge computing for sensor fusion analytics, in which AI calculations used for data fusion will also need to be distributed, employing every usable device on the network. In abstract, this maps to an optimal graph embedding problem (OGEP), with a logical graph embedded within a physical graph. The nodes and edges of the logical graph consist, respectively, of the compute components of distributed AI calculations and the associated communications needed for data collection and transfer. The nodes of the physical graph consist of sensors, available computing resources, and other data source, while communications links act as the physical graph edges. In this scenario, the distribution of tasks performed within a sensor fusion engine could span regions of the physical graph, with (possibly momentarily) fixed sensor and data resources and “agile” compute tasks and communications. In our model, we treat the interactions between nodes in each graph as particles, interacting like atoms in a molecular simulation. Because the available devices will likely have heterogeneous capabilities, we balance logical node placement through use of Coulomb-like forces, with charges determined by resource needs on the logical graph and available resources on the physical graph. We introduce the balanced utilization index (BUI), an adaptation of Jains Fairness Index, to measure this balance. Spring-like forces, based on the physical properties of wireless networking protocols, capture the energy cost of communications. Based on these properties we define an objective function that, when minimized, simultaneously minimizes the communications cost while maximizing the BUI. Although the OGEP is generally intractable, with a highly non-convex solution space, we have developed a simulated annealing method that rapidly and reliably obtains good solutions from the set of local optima of this objective function. We demonstrate this through simulation of an abstract model of a sensor fusion engine, combining data from collocated and distributed sensors, incorporating other analysis data. |