Using a Neural Network as a Function Evaluator during Ga Search for Reliability Optimization

This paper demonstrates the use of a neural network as a reliability estimator for calculation of the objective function value during genetic algorithm (GA) search. The GA searches for the lowest cost system design by selecting the appropriate components and levels of redundancy. Using a neural network approximation for system reliability is computationally efficient for optimization problems where calculation of the objective function is impractical. 1 Corresponding author. INTRODUCTION This paper presents an optimization approach using a genetic algorithm (GA) to identify the preferred choice of components and the optimal levels of redundancy for a reliability design problem. The problem is a combinatorial optimization problem where reliability goals are achieved by discrete choices made from available parts. For complicated design problems, determination of the reliability of a given solution (i.e., system configuration) can require considerable effort or even Monte Carlo simulations. Since GA require numerous objective function evaluations to calculate fitness, a problem where the evaluation of the objective function is computationally time consuming may seem ill-fitted to GA. Our approach to this barrier is to develop a neural network approximation of system reliability. Design of a hardware system involves numerous discrete choices among available components based on cost, reliability, weight, etc. If the objective is to minimize cost for a certain reliability requirement, then a strategy is required to identify the optimal combination of components. This is an NP-hard problem (Chern, 1992). When there are many functionally similar components to choose from, it becomes increasingly difficult to find the optimal solution, particularly when redundancy is considered as a strategy to enhance reliability. Redundancy is the use of functionally similar (but not necessarily identical) components such that if one fails, the redundant component will be available to perform the required function. Figure 1 shows a typical series-parallel system. k-out-of-n redundancy is defined as a series of n parallel components where any k are required to be operating for the system to avoid a failure. The total number of components in parallel, ni, for each function is a variable which is an integer value greater than or equal to ki (for the i subsystem). ki, the required number of components for a given subsystem, is specified while ni remains a variable to be determined through the search.