Genetic algorithm based input selection for a neural network function approximator with applications to SSME health monitoring

Abstract

A genetic ,'dgorithm is used to select the inputs to A neural network function ApproximAtor. lit the application considered, modeling criticM parameters of the Space Shuttle Main Engine (SSME), the functional rel,_tionslfip between mea._ured parameters is unknown and coxuplex. Furthermore, the number of possible input parameters is quite large. MAlty approaches have been used for input selection, but they are either subjective or do not consider the complex multivariate relationships between parameters. Due the optimization altd space searching capabities of genetic Mgorithms they were employed in this paper to systematize the input selection process. The results suggest that the genetic Mgorithm can generate parameter lists of high quality without the explicit use of problem domain knowledge. Suggestions for improving the performance of the input selection process are also provided.

SSME parameter model input selection using genetic algorithms

“…To provide the guidance described above, the fitness function must measure some properties of the candidate solution and compute a number that indicates the "fitness" of that solution. The properties considered in a previous paper include the performance of the generated neural network, the convergence properties of the training process, and the number of inputs [5]. In this work, the soft size constraint used in [5] is replaced by the hard size constraint described above.…”

“…The properties considered in a previous paper include the performance of the generated neural network, the convergence properties of the training process, and the number of inputs [5]. In this work, the soft size constraint used in [5] is replaced by the hard size constraint described above. Furthermore, the convergence properties are not considered since other methods for reducing the effects of training noise are used; these methods are described below.…”

“…Since the variance of the average fitness within a chosen subregion is inversely proportional to the number of samples within that subregion, the use of a large population improves the estimate of the fitness of every subregion in the search space. Here, a population of 250 individuals is used, whereas a population size of 50 was used in [5].…”

“…Unlike steady-state genetic algorithms, where only a single population member is replaced each generation, 50 members of the population were replaced each generation by 50 new chromosomes created through selection and recombination. The number of members to replace each generation was chosen to be 50 in order to reduce selection noise and to facilitate comparison with previous results [5]. Furthermore, to avoid a random search for the first 250 trials, an initial population of 50 chromosomes was initialized as described in Section IIA, and no members of the population were deleted until the population size reached 250.…”

“…These conclusions may be applied to dynamical systems if the system state is sufficiently represented in the inputs of the neural network. For these reasons, feedforward neural networks have been used to model critical parameters of the Space Shuttle Main Engine (SSME) during the start-up transient and they have been shown to be effective [4,5].…”