![]() |
|||
![]()
|
![]() |
![]() |
![]() |
4.2. DISTRIBUTED AGENT METHODSMost configuration-design systems solve the general configuration-design problem, including some that solve the multifunction part and support-function problems, but they do not acknowledge the naturally distributed nature of large-scale configuration-design problems. As a result, they do not take advantage of networks to solve the problem in a distributed way. Each requires a central repository of all design knowledge. In distributed configuration design, the knowledge necessary to solve the problem is distributed among a set of decision-making agents that communicate to achieve shared design goals. These agents are computational processes with:
It is assumed that these agents use their own internal representations and problem-solving methods to contribute to the shared goals. In addition to assuming that design knowledge is distributed, we also assume that the representation of the design is shared among the agents. Thus, there is no shared, global memory, such as a blackboard, that agents read to and write from, and the agents do not send a complete or partial design from agent to agent in some prescribed order. All communication is done by sending messages from one agent to another agent. To solve a configuration-design problem, agents organize themselves into an agent network. Agents are autonomous to the extent that they can reason about their own resources and goals to make a decision whether or not to participate in a design, and they have implicit goal structures. While agents may represent any portion of the design problem, particular kinds of agents are most natural and correspond best to the knowledge in actual design settings. Each type of agent is a decision maker that communicates with other agents. We now identify three classes of agents that solve distributed configuration-design problems (i.e., catalog agents, constraint agents, and system agents) in terms of the features of the design problem they represent, the capabilities they have, and the decisions that they are allowed to make. The design network is a set of agents {catalog-agents, constraint-agents, system-agent} organized to solve distributed configuration-design problems. Catalog agents represent part catalogs. They have the capability to transform their parts into an appropriate representation and to communicate elements of the representation to other agents. Catalog agents make decisions about how to modify attribute domains to achieve design properties defined above, and about which parts to select from a decomposable set to submit as a solution. Thus, the representation of the design is distributed among the catalog agents -- who are the only agents with the authority to modify attribute assignments. Constraint agents represent constraints. Constraint agents have the capability to evaluate an attribute assignment for consistency and decomposability, and to communicate the result of this evaluation to other agents. Constraint agents decide if the assignment is consistent or decomposable, use the constraint-propagation functions to achieve a consistent space, and suggest ways to modify a nondecomposable space, respectively. The system agent represents the required-functions and v() elements of the design problem. A single system agent exists for each instance of a design problem. The system agent has the capability to: (1) create a design network by identifying catalog and constraint agents necessary to solve a configuration-design problem; (2) communicate with other agents; (3) broadcast the problem specifications to the agents in the network; and (4) verify that all functions are implemented by the parts submitted by the catalog agents. The properties defined earlier, that is, consistency and decomposability, apply straightforwardly to catalog and constraint agents. A catalog agent is consistent if and only if (iff) the catalog that it represents is a consistent catalog. A constraint agent is consistent iff the constraint that it represents is consistent. A catalog agent is decomposable iff the catalog that it represents is a decomposable catalog. A constraint agent is decomposable iff the constraint that it represents is decomposable. Properties of the design network follow from the properties of the agents in the network. The design network is a consistent network iff all catalog- and constraint-agents are consistent. The design network is a decomposable network iff catalog and constraint agents are decomposable. A distributed design problem is a tuple <design problem, design network> organized to solve the design problem:
A solution to the distributed design problem is a set of parts S, selected by the catalog agents, with the following properties:
|
![]() |
|
Use of this site is subject certain Terms & Conditions. Copyright (c) 1996-1999 EarthWeb, Inc.. All rights reserved. Reproduction in whole or in part in any form or medium without express written permission of EarthWeb is prohibited. Please read our privacy policy for details. |