Fault-tolerance of a neural network solving the TSP

Summary form only given, as follows. Results are presented of a fault-injection experiment that simulates a neural network solving the traveling salesman problem (TSP). The network is based on a modified version of Hopfield's and Tank's original method. The authors define a performance characteristic for the TSP that allows an overall assessment of the solution quality for different city distributions and problem sizes. Five different 10-, 20-, and 30-city cases are used for the injection of up to 13 simultaneous stuck-at-0 and stuck-at-1 faults. The results of more than 4000 simulation runs show the extreme fault tolerance of the network, especially with respect to stuck-at-0 faults. One possible explanation for the overall surprising result is the redundancy of the problem representation.<
>