Skip to main content

"Large-scale Adaptive Sensing, Learning and Decision Making: Theory and Applications"

Final Report Summary - SCADAPT (Large-scale Adaptive Sensing, Learning and Decision Making: Theory and Applications)

The SCADAPT ERC project addresses one of the fundamental challenges of our time: Acting effectively while facing a deluge of data. It seeks to substantially advance large-scale adaptive decision making under uncertainty, by grounding it in the novel computational framework of adaptive submodular optimization. We are pursuing groundbreaking new scalable techniques, bridging statistical learning, combinatorial optimization, probabilistic inference and decision theory, in order to overcome the limitations of existing methods. In addition to developing theoretically well-founded, rigorous approaches, we are committed to demonstrate the performance of our methods on real-world applications.

Over the course of the project, we have made substantial advances and exciting new discoveries. For example, we have pioneered novel techniques for submodular optimization that can leverage modern computational hardware, by exploiting parallel, distributed and streaming computations. Our algorithms carry strong theoretical performance guarantees, yet outperform classical techniques by several orders of magnitude. We have also discovered novel ways for designing submodular surrogate functions, i.e. objective functions that are adaptive submodular and can hence be highly efficiently optimized, yet still solve the original non-submodular decision problem. We have developed novel computational techniques for succinctly, yet efficiently summarizing massive data sets in a way that is provably sufficient for the purpose of fitting certain statistical models. We have applied our novel methodology in three interdisciplinary problem domains, in community sensing, recommender systems and sequential decision making in computational sustainability.