Layered neural networks, which are a class of models based on neuronal computation in biological systems, are applied to the task of reactivity monitoring in a nuclear reactor to improve the safety and the reliability of the operating plant. Training is done with a maximum likelihood method, which is suitable for on-line training. Operational data from the Fast Breeder Test Reactor are used to study its performance. The adaptability of the network to slow variations in the system parameters and its ability to learn in a noisy environment are studied.