Abstract

Solving sequential decision problems under partial observability is a fundamental but notoriously difficult challenge. I will introduce the new concept of adaptive submodularity, generalizing the classical notion of submodular set functions to adaptive policies. We prove that, if a problem satisfies this property, a simple adaptive greedy algorithm is guaranteed to be competitive with the optimal policy. The concept allows us to recover, generalize, and extend existing results in diverse applications, including sensor management, viral marketing, and active learning. I will focus on two case studies. In an application to Bayesian experimental design in Behavioral Economics, we show how greedy optimization of a novel adaptive submodular criterion outperforms standard myopic techniques based on information gain and value of information. I will also discuss how adaptive submodularity can help to address problems in computational sustainability by presenting results on conservation planning for three rare species in the Pacific Northwest of the United States.

This talk is based on joint work primarily with Daniel Golovin