A Connectionist Network for Dynamic Programming Problems

Dynamic programmingis well-known as a powerful modeling techniquefor dealing with the issue of making optimal decisions sequentially. Many practical problems, such as nding shortest paths in route planning, multi-stage optimal control, can be formulated as special cases of the general sequential decision process. This paper proposes a connectionist network architecture, called the binary relation inference network, which solves a special class of dynamic programming problems in the continuous time. They include the all-pair solutions for a family of closed semiring path problems, such as shortest paths, transitive closure, minimum spanning tree, and minimax path problems. The all-pair inference network speciies a basic and uniform computation of its individual units which then collectively emerge towards a global optimal solution. The computational order in its discrete-timevariants, either as synchronous or asynchronous networks, bear a close resemblanceto the Floyd-Warshall algorithm and doubling algorithm. However, the continuous-time inference network ooers signiicant speed advantage if its non-sequential computation nature can be exploited. Simulation results of using analog VLSI implementation of the inference network for solving shortest path problems are promising.