| Many multiagent systems require allocation of agents to tasks in order to ensure successful task execution. Most systems that perform this allocation assume that the quantity of agents needed for a task is known beforehand. Coalition formation approaches relax this assumption, allowing multiple agents to be dynamically assigned. Unfortunately, many current approaches to coalition formation lack provisions for uncertainty. This prevents application of coalition formation techniques to complex domains, such as real-world robotic systems and agent domains where full state knowledge is not available. Those that do handle uncertainty have no ability to handle dynamic addition or removal of agents from the collective and they constrain the environment to limit the sources of uncertainty. A modeling approach and algorithm for coalition formation is presented that decreases the collective's dependence on knowing agent types. The agent modeling approach enforces stability, allows for arbitrary expansion of the collective, and serves as a basis for calculation of individual coalition payoffs. It explicitly captures uncertainty in agent type and allows uncertainty in coalition value and agent cost, and no agent in the collective is required to perfectly know another agents type. The modeling approach is incorporated into a two part algorithm to generate, evaluate, and join stable coalitions for task execution. A comparison with a prior approach designed to handle uncertainty in agent type shows that the protocol not only provides greater flexibility, but also handles uncertainty on a greater scale. Additional results show the application of the approach to real-world robotics and demonstrate the algorithm's scalability. This provides a framework well suited to decentralized task allocation in general collectives. |
