On-line kernel learning for active sensor networks

In this algorithmic project we consider how to leverage increasingly capable mobile sensor networks to simultaneously exploit a changing environment and maintain low uncertainty about the previously explored regions. This problem, known as ‘coverage-estimation’ or ‘active sensing’ in the literature, has applications in many fields: tracking interesting oceanographic or atmospheric data using a robotic network, finding and clearing environmental hazards such as pollutants [1] or radioactive waste [2], searching for humans in wreckage [3], and even game theoretical formulations such as the multi-armed bandit problem [4]. Much of the literature focuses on either the estimation problem of exploring an unknown environment or the coverage problem of exploiting a known environment.