Multi-Agent Active Search: a Reinforcement Learning Approach

Multi-Agent Active Search (MAAS) is an active learning problem with the objective of locating sparse targets in unknown environment by actively making data-collection decisions. Recently proposed algorithms, although well-motivated from a theoretical perspective, are limited three key ways: they either explicitly myopic (e.g. respect to information gain) or introduce strong biases that fall short fully non-myopic behaviour; employ general-purpose coordination mechanisms scale multi-agent settings without optimising for any specific agent configuration; and involve significant online computation determine suitable sensing regions. In this paper, we Poisson Point Process formulation cast MAAS as Reinforcement Learning problem, policies belief space associated POMDP. We demonstrate how such approach can overcome each issues previous algorithms surprisingly robust test-time miscommunication.