A cyclic pattern resulting from a constraint satisfaction search

A case study has been constructed in the domain of the configuration of telephone exchanges. These issues are considered by using a special symbolic assignment problem including optimization, which was solved in the course of building a knowledge-based system. It is cast into the usual formalism of constraint satisfaction. A unique feature is the absence of an explicit cost function, requiring the use of heuristics for the optimization. The authors treat the results of the problem description analysis in two different approaches: using backtracking and applying a neural net. While the former method is used to solve each problem anew, the net learned patterns for the picking and solving of subproblems. Knowledge about the problem is transformed into static look-ahead and look-back knowledge for the backtrack search. It is used to design the neural net architecture. The authors report on the performance of both approaches. Finally, the authors generalize from the case study, pointing out the unique features not covered appropriately by theoretical work, and present some solutions