Predicting the Behavior of Robotic Swarms in Search and Tag Tasks

Abstract

We are developing an approach for predicting emergent swarm behavior. The swarms consist of agents tasked with tagging targets in simple grid worlds. These swarms are interesting because of their emergent behavior. The path length distributions capture the local probabilistic information. We use networks of simple probabilistic graphs of this local information to model the collective behavior. These networks of local models predict the proportion of agents in each region at a given step. This allows us to predict the effectiveness of tagging in a large grid world given choices of local action by individual agents. We explain the approach, its limitations, and future directions of research.