The paper examines the behaviour of a fuzzy feedforward controller which is trained on-line using a recursive fuzzy identification scheme with local exponential weighting. The control algorithm is derived and its equivalence to conventional proportional-plus-integral control action is established to help in commissioning the controller. The sensitivity of the training algorithm to the distribution of the training data is considered and a modification is proposed that will eliminate training errors. Lastly, experimental results are presented that show that the self-learning controller is able to control the temperature in a liquid helium cryostat which has significant non-linearities.