Abstract
It is shown that a neural net can learn a complex optimization problem such as the traveling salesman problem (TSP) by using contrastive Hebbian learning. Contrastive Hebbian learning is applied to an interactive network to teach the network to solve the TSP from examples. With the use of 'hidden' units, problems of increasing complexity can be learned by a net by increasing the number of hidden units present. The advantages of learning are obvious: one can have the computer design the network, and, once trained, the net will run in constant time. Very successful results were shown for a network trained on several sample problem sets for a four-city TSP.