隨著智慧型手機市場的蓬勃發展，使用者感受(User Experience)正逐漸成為使用者是否購買一隻手機的主要依據。同時，由於直接影響到了手機的回應狀況，應用程式的載入時間(Launch Time)也就自然地成了影響使用者感受的主要因素之一。Android使用了名為Activity Stack的機制來管理應用程式；如果一隻應用程式從Activity Stack內被執行，它的載入時間會較為短暫。因此，如何把「將會被執行的應用程式」保存在Activity Stack內，就成為一個重要的課題。然而，應用程式只有在被使用過才會被放進Activity Stack內；且當Activity Stack滿時，會依據LRU原則而被剔除。理想上，如果我們知道使用者接下來將會執行哪些應用程式，並將這些應用程式保存在Activity Stack內的話，應用程式的載入時間就可以被縮短。所以在這篇文中，我們使用了一種情境感知(Context-Aware)的策略來猜測哪些應用程式將會被執行；我們修改了Android Framework，使系統會在背景自動的預載這些「將會被執行的應用程式」；同時，我們也修改LRU的管理原則，保護這些「將會被執行的應用程式」不被移出Activity Stack。我們紀錄了五個使用者在三個星期內的手機應用程式使用情況，並藉此驗證我們的情境感知管理系統；結果顯示，相比於原生的Android系統，從Activity Stack內被執行的應用程式數量平均上升了27%。

With the burgeoning market of smart phones, user experience is becoming a dominating factor affecting how people buy a smart phone. Though user experience may include a lot of aspects, the launch time of applications definitely influences the perception of the user about the responsiveness of the phone, and thus becomes a significant element of user experience. Android uses a mechanism, Activity Stack, to handle the launch of applications. If an application is launched from the activity stack, the launch time can be shorter. Thus, it is important to keep the applications in the activity stack when they are to be launched. However, applications are loaded into the activity stack on-demand and are removed using the LRU strategy. It turns out that, when an application is "demanded", it is very likely not in the activity stack, leading to a long launch time. Ideally, if we can predict what applications a user will use in the near future and make sure they are loaded and kept in the activity stack, then the application launch time may be reduced. In this thesis, we use a context-aware strategy to predict the future use of applications. We have modified the Android framework to pre-launch an application if it is predicted hot but not in the activity stack. We have also replaced the LRU policy in Android so that a predicted hot application will not be removed from the activity stack. The proposed context-aware prediction method is evaluated based on the activity traces collected from five real users over a three-week period. The results show that our method allows in average 27% more applications to be launched from the activity stack when compared with the original Android system.

[1] An effective approach for mining mobile user habits. CIKM ’10. Cao, Huanhuan and Bao, Tengfei and Yang, Qiang and Chen, Enhong and Tian, Jilei.
[2] An adaptive user interface based on spatiotemporal structure learning. Communications Magazine, IEEE ’11. Lee, Hosub and Choi, Young Sang and Kim, Yeo-Jin.
[3] Aciom: application characteristics-aware disk and network i/o management on android platform. EMSOFT ’11. Kim, Hyosu and Lee, Minsub and Han, Wookhyun and Lee, Kilho and Shin, Insik.
[4] Quality of Service Profiling. ICSE’ 10. Sasa Misailovic, Stelios Sidiroglou, Henry Hoffmann, and Martin Rinard.
[5] An analysis of power consumption in a smartphone. USENIX ’10. Carroll, Aaron and Heiser, Gernot.
[6] On power and throughput tradeoffs of WiFi and Bluetooth in smartphones. INFOCOM ’11. Friedman, Roy and Kogan, Alexander and Krivolapov, Yevgeny.
[7] An Unsupervised Approach to Modeling Personalized Contexts of Mobile Users. ICDM ’10. Tengfei Bao, Happia Cao, Enhong Chen, Jilei Tian, Hui Xiong.
[8] Life cycle of Android applications, http://developer.android.com/reference/android/app/Activity.html
[9] Discovery, analysis, and presentation of strong rules. Knowledge Discovery in Databases ’91. Piatetsky-Shapiro, and Gregory.
[10] Knowledge Discovery in Databases, 1991, Piateski, Gregory and Frawley, William.
[11] Life Cycle of an Android Application http://developer.android.com/reference/android/app/Activity.html
[12] Energy Optimization on Smartphone by Predicting User Behavior Using Hidden Markov Model, 2011, Owen Lo