Considering Agent and Task Openness in Ad Hoc Team Formation

Abstract

When deciding which ad hoc team to join, agents are often re-quired to consider rewards from accomplishing tasks as well as potential benefits from learning when working with others, when solving tasks. We argue that, in order to decide when to learn or when to solve task or both, agents have to consider the existing agents' capabilities and tasks available in the environment, and thus agents have to consider agent and task openness-the rate of new, previously unknown agents (and tasks) that are introduced into the environment. We further assume that agents evolve their capabilities intrinsically through learning by observation or learn-ing by doing when working in a team. Thus, an agent will need to consider which task to do or which team to join would provide the best situation for such learning to occur. In this paper, we develop an auction-based multiagent simulation framework, a mechanism to simulate openness in our environment, and conduct compre-hensive experiments. Our results, based on more than 20,000 sim-ulation runs, show that considering environmental openness is beneficial and necessary, and task selection strategies leveraging openness can improve agent learning and performance. We also report on observations of emergent behaviors related to openness.