This paper deals with an optimal Zero Moment Point (ZMP) based pattern generator for stable dynamic walking in humanoid robotics. The proposed method is based on a Three-Mass Linear Inverted Pendulum Model (3MLIPM), used as a simplified model of the biped robot. The 3MLIPM considers the biped robot as a three point masses and two-link system. A ZMP based performance index is then used in an optimization problem whose solution gives the best values of the model's parameters w.r.t. dynamic walking stability. Numerical simulations are presented to show the effectiveness of the proposed optimal pattern generator for the case of the biped walking robot SHERPA.