Multi-agent Coordination in Planning

  • Authors:
  • Jeroen Valk;Cees Witteveen

  • Affiliations:
  • -;-

  • Venue:
  • PRICAI '02 Proceedings of the 7th Pacific Rim International Conference on Artificial Intelligence: Trends in Artificial Intelligence
  • Year:
  • 2002

Quantified Score

Hi-index 0.00

Visualization

Abstract

We consider coordination problems where several agents, each assigned to some subtask of a complex task, solve their own sub-task by making minimal plans and want to find a common plan based on their individual plans. A task is conceived as a set of primitive tasks (operations), partially ordered by a set of precedence constraints. Operations are distributed among agents dependent on their capabilities and constitute the subtasks the agents have to solve. The precedence constraints between operations in subtasks are inherited from the overall precedence constraints occurring in the task. Since it is assumed that every agent is capable to find a suitable (minimal) plan for its own sub-task, the main problem for the agents to coordinate their plans in order to solve the complete task. First, we characterize situations in which an optimal coordinated plan can be constructed by simple plan coordination. Since, in general, obtaining optimal global plans is intractable, we therefore introduce two simple and efficient distributed approximation algorithms to achieve plan coordination. The first algorithm can be used as a d-approximation of a globally optimal plan for the agents, where d is the depth of the original task, i.e. the length of the longest chain in the set of precedence constraints constituting the task. This algorithm assumes almost no knowledge about the distribution of tasks over the agents. If such knowledge, however, is available, a second, more refined algorithm can be used, that is based on elaborate inter-agent negotiation and is able to achieve a better approximation ratio.