Abstract
This paper investigates low-cost strategies for the multiagent object collection task, in which multiple agents work together to collect a set of items distributed throughout an environment. Several agent architectures are examined, including simple reactive architectures, more complex deliberative architectures, and "predictive" versions of both of these that take other agents into account when choosing targets for collection. A series of "yardstick" experiments demonstrate that the simple agent types perform very well relative to agents that employ much more computationally expensive approaches. Subsequent large-scale simulations that substantially increase the number of agents and collection items demonstrate that both reactive strategies scale well to more realistic task sizes, with the predictive version performing significantly better than the non-predictive ones.