Abstraction Methods for Resource Management in a Distributed Information Network

ion Methods for Resource Management in a Distributed Information Network Berthe Y. Choueiry AI Laboratory Swiss Federal Institute of Technology (EPFL) CH-1015 Lausanne E-mail: [email protected] .ch Dean Allemang R&D, Swiss Telecom CH-3029 Berne E-mail: [email protected] Abstract The complexity of Distributed Information Networks (DIN) becomes prohibitive for large networks with varied and interchangeable services. We claim that this complexity can be managed through the use of appropriate abstractions. In this paper, we describe how abstraction methods can support reactive strategies for the management of resources in a DIN. We rst introduce new results that show the usefulness of abstraction methods for resource allocation. Then, we introduce the problem of resource management in a DIN. Finally, we show how we plan to adapt the methods investigated previously to this new environment. 1 Abstractions for resource allocation It is widely acknowledged that abstractions are useful techniques for reducing the complexity of problem solving. Recently in [2], they have been recognized to be important for supporting intelligent interaction with the users as well. In [2], the authors describe a suite of abstraction methods for resource allocation problems. In this paper, we claim that the following are relevant to planning of Distributed Intelligent Networks (DINs): Decomposition into subproblems and localization of interactions determine parts of the problem between which interactions are limited, Identi cation of interchangeability sets sets of values for certain variables that can be interchanged with no impact on the solution, Isolation of con ict sets small sets of values to which di culties can be localized, Subproblem simpli cation (e.g., aggregation) yields a reduction of problem size, and thus a reduction of the overall problem complexity, and Concept generation exploits background knowledge to build concise descriptions of subproblems and con icts. These techniques can be illustrated with the following example. A resource allocation problem is modeled as a constraint graph; the labels indicate what values may be assigned to each node in the graph, while the edges of the graph indicate that two nodes may not be assigned the same label. The resource allocation problem is to select labels for all nodes that simultaneously satisfy all constraints. Figure 1 (top) shows such a constraint graph. Cluster 2 N4 N5 N1 {g} {g} {g} {c,d,g} N3 N1 Cluster 3 {c,d} {e,f} Cluster 1 Cluster 4 { } N3 N4 N5 { } { } Initial constraint graph