Self Organizing Networks of Gliders for Adaptive Sampling of the Ocean

Underwater Gliders are a cost-effective technological solution to sample the ocean. In highly dynamic areas, adaptive sampling is preferable to static sampling, in order to increase performance by optimally making use of a limited number of gliders. The relatively low cost of these sampling platforms, is opening the door for many research institutions to purchase a fleet of gliders. When a fleet of gliders has to sample for a long period of time in-situ measurement like conductivity, temperature and depth, there is the need to continuously update the glider missions to take into account sampling needs. How to plan the paths of a fleet of gliders is the topic of this work. Using a global cost function, we are able to mesh the area of interest in such a way that finer resolution is used on regions needing a more accurate sampling, while coarser resolution is used on the rest. Once the mesh has been obtained using a meshing algorithm, the associated knots, considered as points in a two-dimensional space, give us an idea on how to adaptively sample the area. This set of points can be fed to a clustering algorithm (like the fuzzy c-means algorithm) in order to get the position of the centroids, where the number of clusters is equal to the number of available gliders. The positions of the centroids can be used as the next surfacing points for each of the associated gliders (a glider is associated to its closest centroid). Since gliders can be programmed to surface approximately at the same instant in time, this control rule allows to self organize the fleet of gliders. At the next surfacing (often programmed to happen in three/four hours), the area of interest is re-meshed, the clustering algorithm is re-executed and the new surfacing positions are re-sent to each glider in order to update their missions. Meteorological and oceanographic (METOC) forecasts of environmental conditions can be used to compute the expected cost function for the next three/four hours. This further improves the performance of the self organizing algorithm. Simulation experiments support the application of this approach. Future work will try to exploit in depth the parallelism between self organizing artificial neural networks (SOANNs) and self organizing glider networks (SOGNs), where gliders play the role of artificial neurons, trying to adapt state-of-the-art learning algorithms developed for SOANN to SOGNs