An Embedded Connectionist Approach for the Inverse Shortest Paths Problem

The Inverse Shortest Path (ISP) problem has recently been considered for applications involving the precise determination of unknown path costs, given only limited experts' knowledge of some shortest paths. Unlike previous approaches which are usually restrictive (such as the minimum change requirement) and algorithmically complicated, a more general but simple optimization scheme based on embedded shortest paths computation is proposed. Conniciting experts' knowledge can also be readily accomodated as multiple objectives instead of hard constraints. In particular, the possibility of embedding a class of connectionist network, called the binary relation inference network, to solve the ISP problem is explored. The inference network has been recently applied in solving constrained optimization problems, such as the shortest path problem, transitive closure, minimax problem, etc. Its inherently parallel operating nature can be well exploited for potential speed-up by embedding it as a real-time engine for the ISP problem. Limitations in incorporating the inference network are discussed and remedies are suggested. Alternatives of using conventional sequential computer and massively parallel machine for the embedded shortest path computation for medium size ISP problems are also investigated.