In this paper, we present a method of training a feedforward neural network using supervised learning scheme to balance an inverted pendulum and cart system. The data used to train the neural network was obtained from a human expert doing the same task. The trained neural network uncovers a set of rules which could be very difficult to derive from the human expert. Comparison was made between the neural-network learned rule and a decision tree rule deducted by Quilan's ID3 induction algorithm using the same set of data. Experiment results showed that the neural network learned rule is more robust. At the same time, we find that the neural network learned rule can be modified to do a similar and more important task——the attitude control of a rocket.