第四代網路將整合各種形形色色的接入技術像是UMTS、WiMAX、WLAN,並引領我們邁入all-IP的網路通訊。在這樣的環境下,多模行動裝置可以在異質的網路間順利地切換,以實現“無所不在”服務的存取。然而如何決策一個最佳的網路給使用者並提供他(她)想要的服務是個重要環節。在過去“總是最佳網路” (Always Best Connected) 的機制是最常被使用的方法;它根據行動終端的RSS、費率等要素輔助其進行存取網路的決策。有別於ABC的方法,近年來一種基於網路連線的“總是最佳連線”(Always Best Network Connection) 被應用於IP語音通訊。不像ABC只考慮發話端的網路選擇,ABNC將發、受話端的網路環境限制及使用者服務需求都納為決策的條件。然而, IP語音只是多媒體服務其中的一部份,在多媒體服務中,有更多其他需求大量的頻寬的媒體是必須考量的。因此,本論文延伸ABNC於多媒體服務中,並善用行動終端之間多條路徑並存的特性來進行網路連線決策。我們將決策問題公式化成“多屬性決策”(Multiple Attributes Decision Making) 與資源配給的問題;並提出創新性的策略來減化決策的複雜度,並以貪婪演算法來減少決策時可挑選的連線組合,最後在幾個情境中的模擬結果證明,我們的決策不但快速,而且有相當好的準確性
The fourth generation (4G) network, which integrates various access technologies, such as UMTS, WiMAX and WLAN, is leading us to an all-IP networking era in the future. In the 4G environment, a handset with multiple interfaces can switch among reachable access networks to achieve ubiquitous service. However, how to select the best access network is critical for a user to obtain his/her preferred service. In the past, ABC (Always Best Connected) has been utilized by a mobile terminal (MT) to select an access network in nearby networks based on receiving signal strength (RSS), cost, etc. Instead of considering access network alone, a new paradigm called connection-based ABNC (Always Best Network Connection) has been presented for VoIP services. ABNC selects an access network based not only on access network but also on properties of communication session, which is composed of both source and destination access networks. Nevertheless, ABNC fails to consider current multimedia services, which requires much larger end-to-end bandwidth than that of VoIP sessions. This thesis extends ABNC to support multimedia services by exploiting parallel paths between multimode MTs. We formulate the problem as a MADM(Multiple Attributes Decision Making) hierarchy with a utility-based media assignment. A series of heuristic approaches and greedy algorithm schemes is proposed to reduce the complexity of access network selection. Finally we evaluate our network selection algorithms with several scenarios. The results illustrate that the performance of the presented algorithms and decision results are closed to optimal.