摘要
拆解式多目標演化演算法 (MOEA/D) 在最近幾年被提出並且在解決多目標最佳化問題上展現了其優異的最佳化能力。它利用參考點 (reference point) 與權重向量 (weight vectors) 將多目標最佳化問題分解成多個子問題後進行最佳化。 在這篇論文中，提出了兩種基於拆解式多目標演化演算法的平行化方法。分解出的子問題被分配到各個不同的執行單元上並且平行化的計算。為了降低各個執行單元溝通所耗費的資源，使用島嶼模型 (island model) 的概念以減少溝通的次數與資訊量。提出了使用多個地區性參考點 (local reference points) 的方式並且將其使用於其中一種平行化方法與原本的拆解式多目標演化演算法中。這篇論文的主要目的是探討提出的平行化方法與使用多個地域性參考點對演算法效能之影響。實驗結果顯示使用多個地區性參考點可以在演化的初期找到比較好的解，但收斂結果不佳。也分析了當參考點為事先給定時，在這些方式中對解品質的影響。該基於拆解式多目標演化演算法的平行化方法可以在較短時間內拿到不錯的解並適用於資訊傳遞量低的環境。
並列摘要
Recently, the multiobjective evolutionary algorithm based on decomposition (MOEA/D) has shown good performance on solving multiobjective optimization problems. The multiobjective optimization problem is decomposed by the combination of reference point and weight vectors. Two parallelization methods of MOEA/D based on the island model are proposed in this study. The population can be distributed to many processors and performed MOEA/D in parallel. The migration method based on the island model reduces the communication cost and the ring network is used in two objective optimization problems. The reference points not only define the search space but also influence the behavior of algorithm. The idea of using local reference points is proposed and adapted to one of the parallelization.The purpose of this thesis is to study the influences of using single and multiple (local) reference points on MOEA/D and the potential benefits of the parallelization methods. The experimental results show that using local reference points can obtain better solution quality in the early generations but can not outperform using single reference point. The method of using static reference point is analyzed. The parallelization can obtain acceptable solution quality with short execution time with very limited communication.