Infotaxis in a Multi-agent Sensor Network
Julia Ebert, Harvard University
Localization of odor and particle emission sources has applications from locating victims in disaster relief to identifying pollution sources. When emissions and detections are sparse, one can no longer rely on traditional gradient-based search techniques such as chemotaxis to locate a source. However, in such sparse conditions insects can still reliably find the locations of food sources or mates by scent. Previous work by Vergassola et al. (2007) demonstrated that with sparse detections, a source can be effectively located by searching according to the information gradient of the system. In this "infotaxis" process, agents move such that they maximize the expected information gain at each point in time. In this work, we extend infotaxis to a multi-agent system, in which multiple virtual agents share their observations of emissions with the goal of one agent locating the source. We find large improvements in search time for even small groups of agents (2-3 agents) and short-distance local communication (range of about 5 percent of the environment). There are also diminishing returns for increasing group size and minimal gain from global over local communication. We observe initial dispersion of the agents to maximize information gain in an uncertain environment (exploration phase) and later clustering as they refine their estimate of the source and can benefit from increased observation density (exploitation phase). This multi-agent approach is analogous to search by social insects and has applications for improving coordinated search in distributed systems.
Abstract Author(s): Julia Ebert, Clark Teeple, Emma Steinhardt, Sharad Ramanathan