Feedforward networks have powerful approximation capabilities without the "explosion of parameters" problem faced by Fourier and polynomial expansions. This paper first introduces feedforward networks and describes their approximation capabilities, then we address several practical issues faced by applications of feedforward networks. First, we demonstrate networks can provide a reasonable estimate of a Bermudagrass hay fertilizer response function with the relatively sparse data often available from experiments. Second, we demonstrate that the estimated network with a practical number of hidden units provides reasonable flexibility. Third, we show how one can constrain feedforward networks to satisfy a priori information without losing their flexible functional form characteristic.