The solutions produced by the multi-objective optimization algorithm are expected to be on the Pareto front which is the shape of the available best solution set at the objective space. As the problem difficulty, objective and decision space dimension is increased, the necessity computational cost for reaching desired convergence and distribution is increased. Two fundamental processes can be applied to reduce the cost. From the first of these, it is succeeded by reducing the computation time at the computation unit. The method used within the thesis is based on increasing the number of computation unit by decreasing the computation time. By this purpose two fundamental models which are master-slave and island models will investigate. At the master-slave model, algorithm is divided into pieces and each of these pieces and they are evaluated at the different computational unit. Since there isn't any big changes at the structure of algorithm and/or problem, solution quality doesn't change. Three processes are evaluated in this model. These processes are i) only objective functions are evaluated at different architectures ii) optimization algorithm is modified to be implemented on different architectures and iii) optimization algorithms are implemented at each of the architectures. Implementations are made both CPU and GPU hardware units. From the obtained results, computation time is decreased but when considered the number of computation unit, it is observed that master-slave model isn't effective. In case of island model, different architectures search different areas of the objective and/or decision space. That's why the performance of the algorithm changes. In thesis, reference point based island model is proposed. This method depends on the distribution of the reference points to different computational units. Although speeding is provided, the linear speed-up in other words same speed decreasing with the number of computation unit isn't supplied. To increase the performance of the proposed method, migration and crossover method is introduced. From the implementations for two objective problems linear speed-up is reached as well as surpasses this speed. To increase the problem difficulty, at the first time in literature hybrid test problems are defined. Same speedup is obtained for these hard problems. Same tests are evaluated for three objective problems. Increasing the number of objective is made difficult to obtain solution. As results of implementations, almost linear speed-up is observed for three objective problems. |