Fault-tolerance in a Boltzmann machine

Connectionist computing models represent a promising advance in the field of neural networks. The Boltzmann machine is a model for connectionist computing that has been applied to the solution of combinatorial optimization problems and which can be viewed as a massively parallel simulated annealing procedure. We have designed a Boltzmann machine to solve quadratic assignment problems, and have demonstrated its effectiveness by comparing its results with optimal solutions, and by comparing its performance with that of other heuristic algorithms. In anticipation of hardware implementation of the Boltzmann machine, it is desirable to develop a quantitative characterization of the inherent fault tolerant properties of this computational model under inevitable conditions of component failure. We have investigated the fault-tolerance of this connectionist model experimentally by injecting a variety of patterns of component failures, including single node failures, column node failures, and random multiple node failures, with the purpose of observing and measuring the deterioration in the quality of the objective function results that are produced. We observed low-percentage degradations in performance that are acceptable from a practical standpoint, and conclude that the Boltzmann model offers an effective and robust heuristic mechanism for combinatorial optimization.