In this paper, a method for dividing the training data as well as training MLP systems to obtain the multiple solutions of IKP of regular manipulators is presented. The sets of training data for each system are strictly defined using the concept of uniqueness domains. The training data are obtained by the forward kinematics and the sign of the determinant of the manipulator Jacobian is used for the determination of the uniqueness domains. An illustrative example with a 3 dof robot with known IK solutions is presented for the verification of the proposed approach.