In this paper we present a new network architecture based on a Multi-Link Neural Network (MLNN) model. An MLNN has a structure similar to that of a feedforward network except that each connection between a pair of nodes in the hidden layer and the output layer is made of multiple links with possibly different weight values. This results in an aggregation of a combinatorial number of subnets that themselves can be viewed as ordinary feedforward networks. For an MLNN with M hidden nodes, N output nodes, and K links for each connection, its overall connection status is represented as an MN dimensional region with KMN sampling points in the weight space. Thus the multi-link structure defines a sampling grid over this region.A region-based search algorithm has been developed that, for problems of a complex nature, offers a better chance in locating a global minimum than the traditional backpropagation algorithm.