Emergent cooperation in multi-agent deliberative planning

Abstract

The author presents preliminary results obtained in the course of developing a research prototype of a multi-agent planning system. He focuses on the planning system's ability to detect, without relying on scripts desired in advance by domain experts. It would place an intolerable burden on domain experts to ask them to specify in advance all patterns of plansteps where cooperation would occur. In order to reason about opportunities for cooperation, the classical concepts of precondition and postcondition are supplemented with the concept of conditions which persist over finite intervals of time. The main preliminary result is that it appears to be significantly simpler and more natural in a multi-agent planner to maintain explicit representations of the intervals of time over which activities persist rather than to represent such activities by discrete time points.<>