應用深層神經網路之替代性多航班動態定價方法

Abstract

本研究嘗試結合動態規劃與深度學習應用在決策與時間有關之問題。動態規劃由於受限於維度詛咒，在求解計算上需耗費相當大量的時間甚至是會導致硬體上的不足而無法計算，另外，現實中有些問題是必須及時處理無法等待的。本研究認為動態規劃的最佳決策中是有規律可循的，若能利用動態規劃的部分決策中找出動態規劃的參數與決策的函數關係，則能克服動態規劃計算上的時間限制，在未來需要重新計算最佳決策時降低計算時間，並應用在求解的規模擴大的情況。 本研究考量一個可替代性航班的機票訂價問題，目標為最大化總銷售金額。我們利用正交實驗設計建立了動態規劃的系統參數組合，並挑選適合的決策作為深層神經網路的訓練資料，並且討論神經網路的架構比較，找出較佳的神經網路模型。最後驗證經過學習後的深層神經網路能夠應用在其他系統參數下，並利用模擬程式比較動態規劃與深層神經網路的決策獲得的銷售額差距。結果表明，本研究所提出的方法可以應用在時間限制的決策問題。

This study attempts to combine dynamic programming and deep learning applied on time-related decision problem. Dynamic programming is limited by the curse of the dimension. It takes a lot of time to solve the problem and even leads to the lack of hardware and cannot be calculated. In addition, some problems in reality must be handled in time and cannot be waited. This study believes that the best decision policy in dynamic programming exit some rule can be functionalization. If we can use the partial decision of dynamic programming to find the function between the system parameters of dynamic programming and corresponding decision, we can overcome the time limit of dynamic programming. This study considers the problem of the ticket pricing for substitutable flights with the goal of maximizing the total revenue. We use the orthogonal experimental design to establish the set of dynamic programming system parameters, and select the appropriate decision as the training data of the deep neural network, and analysis the architecture and hyperparameters of the neural network to find the better neural network model. Finally, we verified that the learned deep neural network can be applied to other system parameters, and the simulation program is used to compare the revenue between the dynamic programming and the deep neural network. The results show that the method proposed in this study can be applied to decision-making problems with time constraints.