Constructing Binary Decision Diagrams using Machine Learning

Binary decision diagrams are a highly popular method for calculating system reliability. By representing the structure function of a binary monotonic system as a binary decision diagram, the calculation of the system's reliability can, in principle, be performed efficiently. However, constructing such diagrams can still be challenging. To ensure that calculations are done quickly, it is important that the diagrams are as compact as possible. In this article, we will show how binary decision diagrams can be constructed using machine learning. The method assumes the existence of a dataset with corresponding values of component states and system states. Such a dataset can easily be generated of any size if the structure function is known and can be calculated efficiently. However, the method can also be used to approximate an unknown structure function based on similar experimental data. The number of possible component states naturally grows exponentially with the number of components in the system. Consequently, if the number of components in the system is high, it will, in practice, not be possible to obtain sufficient data to perfectly describe the system's structure. In the article, we will compare different strategies for handling this problem. Specifically, it is of interest to compare methods that aim to approximate the structure function as accurately as possible with methods that instead focus on estimating system reliability as accurately as possible. The methods will be illustrated with a few examples.

Keywords: Binary decision diagrams, System reliability, Structure functions, Machine learning.