Hierarchies

Organizational hierarchies allow large problems to be solved by agents recursively partitioning them into smaller (perhaps more detailed) problems, which are then handled by other (typically subordinate) agents. This approach has the significant advantage that no single agent needs to grasp the entire problem in all its complexity -- something that is clearly impossible for truly large problems. It has the disadvantage that no single agent grasps the entire problem in all its complexity -- something that can allow organizations to act in truly stupid ways.

Given that the complexity of problems that must be solved by large organizations is beyond the ability of single agents to grasp, however, the idea of breaking problems down hierarchically provides an important potential lever against computational complexity.

CIRL

We are exploring a new approach to hierarchical planning, based on an asynchronous model of communication between planning agents, informed by the way real organizations solve problems. Our approach separates the information processing done by each agent (human or computer) from the information pathways between agents and the information content required on those pathways, and makes it possible to analyze each aspect separately.

One of the main benefits of this approach is that it allows breaking problems down in such a way that the overall solution process remains polynomial. While there must obviously be a cost to this (in terms of some potential solutions that are not obtainable), we conjecture that -- for many problems of interest -- this is not a serious problem. This conjecture is supported by the fact that natural organizations work this way, and in fact often come up with reasonable solutions. Moreover, this approach inherently precludes solutions with certain types of undesireable behaviour, such as extremely high communication overhead.

Important questions include how to determine which agents should handle which tasks, how to merge partial solutions from different agents, how to handle failures of ``subgoals'', and how to determine which agents need which information without swamping the organization and losing the computational benefits of problem partitioning.

Pointers

Parent areas:
Modeling

Implementation areas:
SWFM
O-Plan

Back to the CIRL Overview Page.