This paper explores the efficient construction of the database structure for the human-like arm motion generation using an evolutionary algorithm-based an imitation learning in real-time. The framework of the arm motion generation consists of two processes, imitation learning of human arm motions and generating of a human-like arm motion using the motion database evolved by the learning process in real-time. We aim at constructing the optimized database structure which have the minimum number of one. We compare a human-likeness, similarity of a motion and robot property which minimize a sum of a robot's joint torques for three database structure. We applied our method to the task of teaching a humanoid robot how to make naturally looking movements like catching the cup on the table.