網際服務為時下最熱門的服務導向技術之一，它能藉由各種簡易及標準化的溝通協定來協助跨平台的應用程式整合，然而其分散式及動態連結的特性也引發了許多不同的研究議題，其中以服務品質為基礎的網際服務選擇即是當前最亟待解決的問題之一。由於網際服務的蓬勃發展，越來越多的服務提供者開始發佈各式各樣的網際服務，而使用者也逐漸獲得更多的選擇空間。但針對一個工作流程內的每項功能，若有眾多符合需求的服務可供選擇時，使用者將面臨到該如何依據整體服務品質的效能，從數以萬計的可能中挑選出最佳服務組合的難題。為了解決上述的問題，我們提出一個以服務品質為基礎的網際服務選擇最佳化方法。首先，根據不同結構樣板的特性來定義相對應的服務品質評估方法，並將工作流程設計解構為不同的樣板組合，進而推導出該工作流程的整體服務品質評估標準。其次，援引近似柏拉圖優勢之定義，使用者可依其對不同屬性的主觀敏感度來衡量不同服務組合間的服務品質優劣。根據量化後的使用者主觀敏感度，可將搜尋空間區分為有限個等價類群，輔以所設計的基因演算法逐步地分析、比較，將可找出符合使用者需求的最佳服務組合。我們所提出的最佳化方法有三個主要的貢獻。(1)使用者可自行依主觀偏好設定每一個服務品質目標函數的等價範圍，而搜尋機制將會根據不同使用者的偏好找出相對應的最佳服務組合。因此，使用者可節省過濾主觀認知上相同服務組合的時間。(2)依據不同等價類群在各個服務品質維度的分佈情形，可推導出該次選擇中之最佳服務組合的數目極值，進而可以在搜尋方法上規劃一個適當大小的儲存空間，以確保不會因儲存空間不足而使曾尋獲之最佳服務組合遺失。(3)可利用曼哈頓準則將搜尋空間轉換為度量空間，藉由評估不同服務組合間的品質差距，將可搜尋出大量在各個服務品質維度都具有較好性能的最佳服務組合。 While a lot of services providing identical application functionality are available, Quality-of-Service (QoS) becomes a critical concern to distinguish alternatives from each other. QoS-based services selection is the problem of determining optimal combinations among exponential candidates based on multiple QoS objectives. First, a pattern-wise replacement method is designed to derive composite QoS metrics of a given workflow process. Based on XML-encoded process descriptions and aggregative QoS effects of primitive patterns, it is able to obtain the composite QoS metrics by decomposing non-atomic sub-processes repeatedly. Second, the ε-Pareto dominance relations are employed to help discriminate the QoS quality of combinations in terms of user-defined ε values. Thus the optima can be determined by the set of non-ε-Pareto-dominated combinations. Third, the Epsilon-Pareto Genetic Algorithm (EPGA) is designed to find out the optimal combinations by evolutionary computation. In comparison with related work, the proposed approach has several merits including (1) the ε-Pareto dominance relations enable user to specify subjective sensibility of QoS metrics. Equivalents will be filtered off and only those distinguishable combinations are returned. Therefore, user is saved from identifying equal combinations. (2) According to user-defined ε values for different QoS objectives, it is proved that the number of optimal combinations is finite and estimable. By keeping an archive with appropriate size, all found non-ε-Pareto-dominated combinations will never get lost. (3) The whole search space can be transformed into a metric space based on user-defined ε values and the quality between combinations can be measured by the Manhattan norm. By quantitative quality measurement, the proposed approach enables more explorations of optimal combinations that have more prominent performance in every QoS dimension than others.
