在5G/B5G網路中使用切片方式進行不同服務的支援是5G/B5G網路的關鍵技術。已有許多具有彈性的控制方法應用在允入控制及資源管理等問題上，這些控制方法的共同特性是它們均使用目前網路環境的參數，例如網路最大及最小的傳輸速率以及可用資源。 然而，由於網路切片使用虛擬技術共享實體資源，使得網路切片之資源管理益趨複雜，因此，如何隨著資源的使用狀況進行切片的管理是非常困難的。為了解決這個問題，在本論文中我們提出了一深度強化學習演算法及深度神經網路的管理系統，用以預估未來時間內資源的使用狀況，並經由深度神經網路策略選取適合的切片請求，藉由與環境的互動達到獎勵的最大化。在切片請求之管理策略上，我們提出分階段的切片管理策略，先使用監督學習來進行參數學習，再透過與環境互動的強化學習更新參數，最後進行切片請求的管理。透過不同階段的資料收集進行再訓練，持續更新策略網路參數，獲得更好的控制結果。除此之外，本論文採用了兩種綠色學習方法，使用預訓練以及減少模型參數的數量來加快模型訓練達到更好的效能。;Using slicing to support different services in 5G/B5G networks is a key technology. Many flexible control methods have been applied to issues such as admission control and resource management. The common feature of these control methods is that they all use parameters of the current network environment, such as the maximum and minimum transmission rates of the network and available resources. However, since network slicing uses virtual technology to share physical resources, resource management of network slicing becomes increasingly complex. Therefore, it is very difficult to manage slices according to resource usage. In order to solve this problem, in this paper we propose a deep reinforcement learning algorithm and a deep neural network management system to predict the usage of resources in the future. Network slice requests are accpted through deep neural network strategies, and rewards are maximized through interaction with the environment. We propose a staged slice management strategy, which first uses supervised learning for parameter learning, then updates parameters through reinforcement learning that interacts with the environment, and finally manages slice requests. Retraining is performed through data collection at different stages, and policy network parameters are continuously updated to obtain better control results. In addition, this paper adopts two green learning methods, using pre-training and reducing the number of model parameters to speed up model training and achieve better performance.
