• A multi-objective optimization method using game algorithm for parallel manipulators is proposed. • A new stiffness index considering the coupling effect of the non-diagonal elements is proposed. • A new index which defines the difference between the total and principal diagonal stiffness indices is proposed. • A four-dimensional intuitive image of slice distribution of the index is presented. In this paper, a multi-objective optimization game algorithm (MOOGA) for parallel manipulators (PMs) is proposed. The proposed algorithm considers the volume of the regular cylindrical workspace, motion/force transmission performance, and stiffness performance as objective functions. First, the distributions of the objective functions in the complete parameter space are calculated and sorted by importance. Second, game weighting factors and lower bound values are assigned to different objective functions according to the engineering requirements. Finally, after multiple rounds of gaming according to the weighting factors and lower bound values, the objective functions reach an optimal balance point and obtain a balance intersection subspace. In addition, a new comprehensive stiffness index (CSI) is proposed, that takes the coupling of the non-diagonal elements into consideration. This index decouples the linear and angular stiffness and has definite physical dimensions as well as a clear physical meaning. A Lagrangian function is used to obtain the maximum and minimum stiffnesses at a given position along with their corresponding directions. To compare the difference between the CSI and the principal diagonal stiffness index (PDSI), a divergence index κ is proposed. 2UPR–RPU and 2UPR–2RPU PMs are employed as examples to implement the proposed algorithm, where U, P and R denote a universal joint, prismatic pair and revolute pair, respectively. The corresponding slice distributions of the local CSI and κ in the regular cylindrical workspace are presented. Additionally, the distributions of the extreme linear stiffness indices and their corresponding directions are presented. The results show that the CSI is decreased by 99% relative to the PDSI. The numerical results demonstrate the effectiveness of the algorithm proposed in this paper.