A quasi optimal solution for the linear fixedend-point, fixed-time problem can be realized by repeated trial and correction of the compensation input to the state-feedback regulator when the knowledge of the system dynamics is inaccurate. The paper presents a method of modifying the input in order to reduce the terminal error using the one obtained in each trial, and shows the convergence condition of this learning process. It is also shown that a similar treatment can be done for the system output which has the same dimensions as those of a model when the dimensions of the object system are greater than those of the model. Simulation and experimental results are given for the settling control of a bridge-crane model.