With the rapid accumulation of gene expression data in publicly accessible databases, computational study of gene regulation has become an obtainable goal Intrinsic to this task will be data mining tools for inferring knowledge from biological data. In this project, we have developed a new data mining technique in which we adapt the connectivity of a recurrent neural network model by indexing regulatory elements and including nonlinear interaction terms. The new technique reduces the number of parameters by O(n), therefore increasing the chance of recovering the underlying regulatory network. In order to fit the model from data, we have developed a genetic fitting algorithm with O(n) time complexity and that adapts the connectivity during the fitting process until a satisfactory fit is obtained. We have implemented this fitting algorithm and applied it to two data sets: rat central nervous system development (CNS) data with 112 genes, and yeast whole genome data with 2467 genes. With multiple runs of the fitting algorithm, we were able to efficiently generate a statistical pattern of the model parameters from the data. Because of its adaptive features, this method will be especially useful for reconstructing coarse-grained gene regulatory network from large scale or genome scale gene expression data sets.