Experimenting Abstraction Mechanisms Through an Agent-Based Hierarchical Planner

In this paper, an agent-based architecture devised to perform experiments on hierarchical planning is described. The planning activity results from the interaction of a community of agents, some of them being explicitly devoted to embed one or more existing planners. The proposed architecture allows to exploit the characteristics of any external planner, under the hypothesis that a suitable wrapper –in form of planning agent– is provided. An implementation of the architecture, able to embed one planner of the graphplan family, has been used to directly assess whether or not abstraction mechanisms can help to reduce the time complexity of the search on specific domains. Some preliminary experiments are reported, focusing on problems taken from the AIPS 2002, 2000 and 1998 planning competitions. Comparative results, obtained by assessing the performances of the selected planner (used first in a stand-alone configuration and then embedded into the proposed multi-agent architecture), put into evidence that abstraction may significantly speed up the search.