An algorithm-structured neural net for the shortest-path problem
Gerald Fahner · 2002
A neural-net-inspired non-von-Neumann architecture which implements Dijkstra's (1959) dynamic programming algorithm to find the shortest path between two given nodes in a graph is presented. The net consists of two layers of binary higher-order neurons, where each layer is fully connected. The first layer neurons act as a recurrent net whose dynamics branches through the search tree. To guide the search, this layer is supervised by additional units that handle real-valued data involved in the search. During the search, the first layer feeds the second layer, which acts as an optimal policy table. This layer is supervised to record relevant information about the intermediate stages of the search. After the search has terminated, the first layer is no longer used. The second layer, working recurrently, outputs the optimal sequence of nodes.>