Neural Nets for Network Reliability

By network reliability, we mean the probability of survival of a network for a speciÞed period of time. Obtaining this probability boils down to the formulation and the evaluation of the reliability polynomial. This is a formidable task for networks with a large number of nodes, and an almost impossible task if the node life-times cascade and interact. Consequently, implementing a mathematical theory of network reliability for even the simplest networks encountered in practice is fraught with obstacles. Neural network based methodologies have recently witnessed a surge of applications. In this paper, we investigate if such methodologies could be a way to circumvent the obstacles mentioned before, albeit, by way of an approximation. For this, we consider a simple bridge structure network with dependencies assumed for the life-times of some of its nodes, and compute its exact reliability. We then simulate failure data from this network incorporating the assumed dependencies, and use these data to train a neural net. The trained neural net is then used to estimate the reliability of the bridge structure. Based on a comparison of the exact and the neural net based reliabilities, we claim that network reliability assessment based on a neural net is a viable approach, but only when a plethora of data is available, for training the net.