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