A novel on-line optimizing control strategy for wall-cooled fixed-bed reactors is presented. It is essentially a nonlinear model predictive algorithm in which the profit is uniquely expressed as the objective function, instead of a square sum of the deviations between model predicted outputs and a desired output variable trajectory. Excessive bed temperature is avoided by setting a constraint on the maximum reactor temperature. Excellent performance including fast tracking to the optimum operating condition and small overshooting of the reactor temperature has been demonstrated by taking advantage of a dynamic KL-NN reactor model.