Neural networks have been shown to be capable of predicting the moderator temperature coefficient in a nuc1ear reactor by using the frequency response functions between the in-core neutron flux signal and the ex-core thermocouple signal as inputs. In this work, actual data from a nuc1ear reactor is used by neural networks to estimate the moderator temperature coefficient at different times during a fuel cycle. Along with the conventional method of training neural networks, a new method of training that better models the use of neural networks in predicting the moderator temperature coefficient is also successfully demonstrated. The results show that neural networks are effective at estimating the moderator temperature coefficient if the domain of prediction is within the training domain of the network. The advantage of using the autoregression method to create the frequency response patterns used as inputs to the neural network as opposed to frequency response functions calculated by the Fourier transform method is also shown. / Master of Science